© The Authors 2025

1 Introduction

Comparing observations of young exoplanets with those of older counterparts provides constraints on theoretical models of planet formation and evolution. However, very young stars are often active and rapidly rotating, rendering their close-in planets challenging to detect and characterize through the usual photometric technique from dedicated spacecraft (for planets transiting in front of their host stars), as well as with the velocimetric method involving ground based, ultra-stable, high-resolution spectrometers. For this reason, only a handful of planets around stars younger than 20 Myr have been reliably confirmed so far, implying that we still have very few observational constraints on the progenitors of planets found around mature stars.

AU Mic, a member of the β Pic moving group (aged ≃20 Myr, Mamajek & Bell 2014; Miret-Roig et al. 2020), is a key object in this respect, being one of the closest and brightest pre-main-sequence (PMS) M dwarfs. Known to host an extended debris disc with moving features (Kalas et al. 2004; Boccaletti et al. 2015, 2018) and two transiting warm Neptunes (Plavchan et al. 2020; Martioli et al. 2021), AU Mic is an ideal target for studying the formation and evolution of young planets and their atmospheres (Hirano et al. 2020). It has been extensively monitored with both photometry and spectroscopy (Klein et al. 2021; Cale et al. 2021; Klein et al. 2022; Zicher et al. 2022; Szabó et al. 2022; Donati et al. 2023a; Wittrock et al. 2023; Mallorquín et al. 2024; Yu et al. 2025; Boldog et al. 2025) to further characterize the two transiting planets (AU Mic b and c) and to search for additional ones in the system. In addition, AU Mic is also known for its intense activity and strong magnetic field (Kochukhov & Reiners 2020; Klein et al. 2021; Donati et al. 2023a), making it a prime target for studying dynamos of mostly convective stars, magnetized winds and star-planet interactions (Kavanagh et al. 2021; Klein et al. 2022; Alvarado-Gómez et al. 2022), as well as for escaping planetary atmospheres (Hirano et al. 2020; Carolan et al. 2020; Allart et al. 2023; Masson et al. 2024).

We carried out a multi-season monitoring campaign of AU Mic from early 2019 to mid-2022 (Donati et al. 2023a) with the SPIRou nIR spectropolarimeter / high-precision velocimeter (Donati et al. 2020) at the Canada-France-Hawaii Telescope (CFHT), mostly within the SPIRou Legacy Survey (SLS), a Large Programme of 310 nights with SPIRou studying planetary systems around nearby M dwarfs and planet formation around young magnetic stars. This monitoring allowed us to document the year-to-year evolution of the large- and small-scale magnetic field of AU Mic. It also enabled us to further characterize its planetary system, with the detection of a candidate outer planet at a distance of 0.17 au the existence of which was recently challenged by new observations in the optical domain, where activity is much stronger (Mallorquín et al. 2024).

In this paper, we present follow-up monitoring of AU Mic from mid-2022 to late 2024 with SPIRou. This work was first carried out within SPICE, a Large Programme of 174 nights at CFHT aiming to consolidate and enhance the results of the SLS until mid-2024, then extended with a PI programme in semester 2024B (PI: J.-F Donati, runIDs: 24BF15 and 24BF96), for an overall time coverage of 2041 d (5.6 yr). We outline these new observations in Sect. 2 and describe the modeling of the time series using Gaussian process regression (GPR) in Sect. 3. In Sect. 4, we model all spectropolarimetric data of AU Mic with Zeeman-Doppler imaging (ZDI) in a homogeneous way. We perform a new velocimetric analysis of the radial velocity (RV) data collected since the beginning of this monitoring effort in Sect. 5. An overview of activity indexes is presented in Sect. 6. We summarize the results in Sect. 7 and discuss their implications on the characterization of AU Mic and its planetary system as well as the more general understanding of star and planet formation.

2 New SPIRou observations

During this monitoring program, we recorded 161 new spectra of AU Mic with SPIRou: 18 in late 2022, 76 in 2023 and 67 in 2024. For four of them (on 2023 May 03, and 2024 Nov 19, 23 and 25), poor and / or irregular weather yielded lower quality spectra with signal-to-noise ratios (S/Ns) per 2.3 km s⁻¹ pixel below 200, which we discarded from the analysis. This yielded a total of 157 spectra (18,75 and 64 in late 2022, 2023, and 2024, respectively) with S/N ranging from 308 to 954 (median 825). We added these 157 new spectra to the 225 older ones collected in the previous monitoring (Donati et al. 2023a), yielding a total of 382 spectra covering a period of 2041 d.

SPIRou records spectra covering the entire 0.95–2.50 μm (YJHK) wavelength range in a single exposure, at a resolving power of 70 000 (Donati et al. 2020). We obtained all spectra of AU Mic in circular polarization (Stokes V) mode, with SPIRou polarization sequences consisting of four sub-exposures, each associated with a different orientation of the Fresnel rhomb retarders (to remove systematics in polarization spectra to first order, see Donati et al. 1997). Each recorded sequence yielded one Stokes I and one Stokes V spectrum, as well as one null polarization check (called N) used to diagnose potential instrumental or data reduction issues. In cases of unstable weather where only two sub-exposures were recorded, we could still retrieve a pair of Stokes I and V spectra, but no N spectrum was obtained and the compensation of systematics was slightly worse. Total exposure times per visit ranged from 379 to 802 s (median 802 s).

All data were homogeneously processed with Libre ESpRIT, the nominal reduction pipeline of ESPaDOnS at CFHT adapted for SPIRou (Donati et al. 2020). We used these reduced spectra in particular for the spectropolarimetric analyses outlined in Sects. 3 and 4. We applied least-squares deconvolution (LSD, Donati et al. 1997) to all reduced spectra, using a line mask constructed from the VALD-3 atomic and molecular line database (Ryabchikova et al. 2015) for an effective temperature T_eff=3750 K and a logarithmic surface gravity log g=4.5 adapted to AU Mic. We only selected the ≃1500 atomic lines deeper than 10% of the continuum level, yielding an average wavelength and Landé factor of 1750 nm and 1.2, respectively. The noise levels σ_V in the resulting Stokes V LSD profiles ranged from 0.74 to 2.16 (median 0.95) in units of 10⁻⁴I_c, where I_c is the continuum intensity.

From these reduced spectra, we derived a homogeneous set of Stokes I and V LSD profiles and corresponding measurements of the longitudinal field B_ℓ, i.e., the line-of-sight-projected component of the vector magnetic field averaged over the visible hemisphere, following Donati et al. (1997). As the Stokes V LSD signatures of AU Mic are quite broad, we computed the first moment over a domain of ±45 km s⁻¹ about the line center, and estimated the equivalent width of the Stokes I LSD profiles through a Gaussian fit (yielding ≃2 km s⁻¹). The reduced chisquare ${\chi }_{\mathrm{r}}^2$ of the B_ℓ data (with respect to B_ℓ = 0) is equal to 401, implying that the large-scale field of AU Mic is unambiguously detected. We also calculated LSD profiles for N, which yielded ${\chi }_{\mathrm{r}}^2$ = 1.1, consistent with no spurious signal down to the noise level and indicating that there are no issues in the observation and reduction procedures, or in the derivation of B_ℓ and corresponding error bars. The inferred B_ℓ values range from −239 to 257 G (median 39 G) over the 2041 d of the monitoring period, with error bars from 3.6 to 10.8 G (median 4.7 G). The corresponding log of Libre ESpRIT reductions and associated quantities is given in Table A.1, with rotation cycles and phases computed with a rotation period of 4.86 d and an arbitrary reference barycentric Julian date of BJD0 = 2 459 000.

We also processed AU Mic observations with APERO (v0.7.292), the latest version of the SPIRou reduction pipeline (Cook et al. 2022), much better optimized in terms of RV precision than Libre ESpRIT. We then analyzed the reduced spectra using the line-by-line (LBL) technique (v0.65, Artigau et al. 2022), with the median APERO spectrum of AU Mic itself as the reference. We computed precise RVs for the 344 nightly-averaged observations, with 154 new RVs added to the 190 older ones from the previous monitoring (and three of the 157 new spectra discarded by APERO). We corrected these RVs for spectrograph drifts using the Fabry-Perot spectrum that SPIRou records simultaneously with the stellar spectrum (Donati et al. 2020). The RV changes of AU Mic (with respect to its average RV of −4.51 km s⁻¹) ranged from −120 to 95 m s⁻¹ throughout the monitoring period, with nightly averaged error bars spanning from 0.6 to 5.7 m s⁻¹ (1.5–3.7 m s⁻¹ excluding the transit night of 2019 June 16 and the two worst quality RV points, median 2.4 m s⁻¹). From the variation in the depths of spectral lines, LBL also derives a precise estimate of the change in effective temperature dT over the visible stellar hemisphere (Artigau et al. 2024), as a result of starspots appearing and disappearing as the star rotates and the surface brightness distribution evolves with time. We found that dT ranged from −20.2 to 25.5 K, and the error bars from 0.4 to 3.0 K (median 1.3 K).

As in Donati et al. (2023a), we also analyzed the nightly averaged spectra with ZeeTurbo to estimate the small-scale magnetic field <B> at the surface of AU Mic and its temporal evolution (following Cristofari et al. 2023), with all nonmagnetic stellar parameters fixed to the values derived in the previous study from the median spectrum of AU Mic (see Table 1 of Donati et al. 2023a). We found that <B> ranged from 2.31 to 3.08 kG (median 2.74 kG) with error bars of 0.030–0.065 kG (median 0.038 kG). The log of APERO reductions and associated measurements is given in Table A.2.

3 Magnetic field & temperature changes of AU Mic

Following Donati et al. (2023a), we employed the framework of Haywood et al. (2014) and Rajpaul et al. (2015) to carry out a quasiperiodic (QP) GPR fit to the B_ℓ values, arranged in a vector denoted y. The QP covariance function c(t, t′) that we used for this purpose is as follows: $$c(t,t^{\mathrm{\prime }})={\theta }_1^2\exp (-\frac{{(t-t^{\mathrm{\prime }})}^2}{2{\theta }_3^2}-\frac{{\sin }^2\left(\frac{\pi (t-t^{\mathrm{\prime }})}{{\theta }_2}\right)}{2{\theta }_4^2})$$(1)

where θ₁ is the amplitude (in G) of the Gaussian process (GP), θ₂ its recurrence period (measuring P_rot), θ₃ the evolution timescale on which the B_ℓ curve changes shape (in d), and θ₄ is a smoothing parameter describing the amount of allowed harmonic complexity. We added a fifth hyperparameter, θ₅, that describes the excess uncorrelated noise needed to obtain the QP GPR fit to the B_ℓ data with the highest likelihood ℒ, defined by $$2\log \mathcal{L}=-n\log (2\pi )-\log |C+\mathrm{\Sigma }+S|-y^T(C+\mathrm{\Sigma }+S{)}^{-1}y$$(2)

where C is the covariance matrix for all observing epochs, Σ is the diagonal variance matrix associated with $y,S={\theta }_5^{2}J$ is the contribution of the additional white noise, with J the identity matrix, and n the number of data points. We then used a Monte-Carlo Markov chain (MCMC) process to explore the hyperparameter domain, yielding posterior distributions and error bars for each. We used the same MCMC and GPR modeling tools as in previous studies (e.g., Donati et al. 2023a,b, 2024b,a). The MCMC process used a conventional single chain Metropolis-Hastings scheme, typically running over a few 10⁵ steps, including the first few 10⁴ steps as burn-in. Convergence was checked with an autocorrelation analysis, verifying that the burn-in and main phase exceed the autocorrelation lengths of all parameters by more than a factor of 10. We computed the marginal logarithmic likelihood log ℒ_M of a given solution following Chib & Jeliazkov (2001), as in Haywood et al. (2014).

The result of the fit is shown in Fig. 1, with a zoom on the 2023 and 2024 data also provided in Figs. B.1 and B.2 (see Donati et al. 2023a, for a zoom on the 2020 and 2021 data). The GP parameters and error bars from the fit are listed in the top section of Table B.1. All parameters are well defined, notably the recurrence period θ₂, which we find to be 4.8591 ± 0.0019 d (consistent within the error bar with the estimate of Donati et al. 2023a) and the evolution timescale, 101 ± 9 d, i.e., half the duration of a typical observing season. The data were fit to an rms level of 7.2 G, larger than the average error bar on the B_ℓ measurements (4.7 G), yielding ${\chi }_{\mathrm{r}}^2$ = 2.3. The B_ℓ data were thus not fit down to the photon-noise level, suggesting an additional source of noise, such as intrinsic variability caused by activity (e.g., stochastic changes in the large-scale field, flares), which was modeled by GPR with θ₅ > 0.

The seasonal variations of B_ℓ are clearly detected, with the modulation amplitude shrinking to a minimum in October 2019 (BJD 2 458 800), reaching a maximum in the next season (BJD 2 459 100), then shrinking again to a minimum in 2023 (BJD 2 460 100). Moreover, as expected from the rather short evolution timescale, the B_ℓ curve from the fit also evolves significantly within each season (see Fig. B.1). Running GPR on individual seasons with dense coverage further suggests that evolution was faster in 2020 (θ₃ = 73 ± 10 d) than in 2021 (θ₃ = 140 ± 45 d) and 2023 (θ₃ = 160 ± 45 d). The rotation period also fluctuates slightly from 4.8536 ± 0.044 d (in 2020) up to 4.8662 ± 0.0045 d (i.e., by about 3σ), likely as a result of differential rotation.

As for B_ℓ, we carried out a QP GPR of the small-scale field <B>, yielding the results shown in the middle panel of Fig. 1 and the hyperparameters listed in the middle section of Table B.1, with ${\chi }_{\mathrm{r}}^2$ = 0.98. We find that <B> is also modulated by the rotation cycle, with a recurrence period of 4.8633 ± 0.0016 d, consistent at the 2σ level with the period derived from B_ℓ. The semi-amplitude of the <B> modulation is θ₁ = 0.135 ± 0.016 kG on average, reaching a minimum of 0.1 kG in mid-2020 and a maximum of 0.3 kG in 2022. We also note that <B> poorly (anti)correlates with B_ℓ (Pearson’s coefficient R=−0.4) and even less with |B_ℓ| (R=−0.2). In fact, the modulation of <B> is smallest when that of B_ℓ is largest (in 2020) and vice versa (in 2023). This behavior reflects that B_ℓ is very sensitive to the orientation of the magnetic field, whereas <B> is virtually not, yielding significantly different modulation patterns. We also report that the evolution timescale is much longer for <B> than for B_ℓ, by a factor of about two, partly reflecting that the temporal evolution of <B> is of much lower amplitude (with respect to the measurement error bar) and therefore less precisely characterized than that of B_ℓ. Finally, we note that <B> decreases from an average of 2.81 kG at the beginning of the observations, to about 2.64 kG in 2023, before increasing slightly (to 2.68 kG) in the last season.

We also carried out the same analysis for the temperature change dT of AU Mic, yielding the results shown in the bottom panel of Fig. 1 and the hyperparameters listed in the bottom section of Table B.1. We find in particular that dT is rotationally modulated, with a period of 4.8611 ± 0.0015 d, again consistent at the 2σ level with the periods obtained from both B_ℓ and <B>. The semi-amplitude in the change of effective temperature (resulting from surface spots appearing and disappearing as the star rotates) is equal to θ₁ = 10.5 ± 1.2 K on average, varying from ≃5 K in 2020 up to almost 25 K in 2022. The evolution timescale is 1.6× longer than for B_ℓ and better defined than for <B>, reaching 167 ± 11 d, demonstrating that topological changes in the large-scale field occur faster than changes in the location, intensity or contrast of the small-scale field and associated surface brightness features. Moreover, we see that, as already noted in Artigau et al. (2024) in a smaller subset, dT mimics <B> quite closely, the latter being strongly anticorrelated with the former at a level of R=−0.92. This implies that dT is a reliable proxy for <B> in the particular case of AU Mic (and for a few other M dwarfs; Cristofari et al. 2025). The measured change rate between <B> and dT is −76 K/kG, with an intercept of +210 K for no magnetic field, suggesting that AU Mic, if unspotted and nonmagnetic, would have a temperature about 210 K higher, i.e., T_eff ≃ 3875 K. Assuming a typical spot-to-photosphere temperature contrast of 620 K for AU Mic (Berdyugina 2005), we infer that the average fraction of the star covered with spots is 210/620=34%. The average dT is minimum at the beginning of the observations (−5.5 K) and rises until the penultimate season (reaching 5.5 K), then decreases in 2024 (to 2.5 K), implying seasonal changes in the spot coverage of AU Mic similar to those induced by rotational modulation, i.e., ±1–2%.

4 Zeeman-Doppler imaging of AU Mic

For this study, we performed complete reanalysis of the Stokes IV data sets of AU Mic over the full timescale of the monitoring. We started by splitting data from each season into two subsets more or less corresponding to A and B semesters (except for 2019, for which we only collected enough data for a dense phase coverage in semester B), yielding a total of 11 subsets (from 2019 B to 2024 B) spanning between four and 25 rotation cycles of AU Mic (i.e., 20 to 120 d; median 70 d). Some of the 2023 B data are also included in another magnetometric study of AU Mic, which exploits additional Stokes QU observations (Donati et al. 2025a).

We used the same ZDI code as in previous analyses (e.g., Donati et al. 2023a), which allows reconstruction of the relative photospheric brightness distribution and the topology of the large-scale magnetic field at the surface of a rotating star from phase-resolved sets of Stokes I and V LSD profiles. This is achieved through an iterative process, starting from a small magnetic field and a featureless brightness map and progressively adding information to the surface of the star, exploring the parameter space until the modeled Stokes I and V profiles match the observed profiles at the required level, typically ${\chi }_{\mathrm{r}}^2$ ≃ 1 (see, e.g., Brown et al. 1991; Donati & Brown 1997; Donati et al. 2006, for more information on ZDI). In practice, the stellar surface is described as a grid of 5000 cells. Whereas the relative photospheric brightness is simply described as a series of independent pixels, the large-scale magnetic field is expressed as a spherical harmonics expansion, using the formalism of Donati et al. (2006, see also Lehmann & Donati 2022; Finociety & Donati 2022; Donati et al. 2023a), in which the poloidal and toroidal components of the vector field depend on three sets of complex coefficients: α_ℓ,m and β_ℓ,m for the poloidal component and γ_ℓ,m for the toroidal component. Here, ℓ and m denote the degree and order of the corresponding spherical harmonic term in the expansion. As in Donati et al. (2025a), we used a spherical harmonic expansion with terms up to ℓ = 10, which is sufficient given the moderate rotational broadening of spectral lines (v sin i = 8.5 ± 0.2 km s⁻¹; Donati et al. 2023a) in the spectrum of AU Mic. As this inversion problem is ill-posed, ensuring a unique solution requires regularization, which in the present case follows the principles of maximum entropy image reconstruction (Skilling & Bryan 1984) to select the image with minimal information among those matching the data. In these ZDI reconstructions, we again assumed an inclination angle i = 80° between the rotation axis and the line of sight, i.e., slightly lower than the inclination of the orbital planes of planets b and c (e.g., Szabó et al. 2022), to reduce mirroring effects of the imaging process between the upper and lower hemispheres.

To compute local synthetic Stokes IV profiles from each grid cell, we used Unno-Rachkovsky’s analytical solution of the polarized radiative transfer equation in a plane-parallel Milne-Eddington atmosphere (Landi degl’Innocenti & Landolfi 2004). We then integrated the spectral contributions from all visible grid cells, assuming a linear center-to-limb darkening law for the continuum, with a coefficient of 0.3, to obtain the global synthetic profiles at each observed rotation phase. The mean characteristics of the synthetic profiles were copied from those of the observed LSD profiles, i.e., a central wavelength of 1700 nm, a Landé factor of 1.2, and a Doppler width of v_D = 3.5 km s⁻¹ (Donati et al. 2023a). We introduced a first filling factor, f_V, (constant over the whole star) describing the fraction of each grid cell that contributes to the large-scale field and to Stokes V profiles, implying a magnetic field of B_V / f_V in the magnetic portion of the cell and a magnetic flux of B_V over the whole cell. Similarly, we assumed that a fraction f_I of each grid cell (the filling factor of the small-scale field, again equal for all cells) hosts small-scale fields of strength B_V / f_V, implying a small-scale magnetic flux over the whole cell of B_I = B_V f_I/f_V. In this context, <B> measured with ZeeTurbo at a given epoch equals the weighted limb-darkened average of B_I over the visible stellar hemisphere. This simple approach within ZDI, used to empirically reproduce the coexistence of small-scale and large-scale fields, was found to provide an adequate description of the Stokes IV data sets of AU Mic (Donati et al. 2023a, 2025a). In this paper, we chose not to apply ZDI to Stokes V profiles only, given the strong underestimation of both the large- and small-scale field that this approach leads to Donati et al. (2023a). We also chose not to attempt estimating surface differential rotation, which occurs on a timescale comparable to the evolution of the large-scale field and is thus hard to reliably detect with ZDI (Donati et al. 2023a).

We applied ZDI to the 11 data subsets and show the achieved fits to the LSD Stokes I and V profiles in Fig. C.1, as well as the recovered magnetic maps in Fig. 2 for the last 4 subsets (2023A to 2024B). The reconstructed maps from previous semesters are shown in Figs. C.2 and C.3. Photospheric brightness maps were also reconstructed simultaneously as part of the imaging process but show only a few low contrast features and are therefore not shown in Fig. 2. On stars with low v sin i, ZDI reconstructs mostly the non-axisymmetric component of the brightness distribution, inducing the observed modulation of Stokes I LSD profiles and covering only a few % of the stellar surface (in agreement with the observed modulation of dT; see Sect. 3). As described in the previous study, fitting the LSD Stokes I and V profiles simultaneously allows one to obtain a more reliable description of both the large-scale and small-scale fields of AU Mic, compared to a reconstruction using Stokes V data only, which suffers more cancellation from the northern and southern hemisphere given the almost equator-on orientation of AU Mic. The epoch-to-epoch variations are real, as is evident from the B_ℓ and <B> times series (see Sect. 3 and Fig. 1, although not so obvious to diagnose visually from the maps themselves. We list in Table C.1 the quantitative characteristics of the reconstructed magnetic topologies for each of the 11 subsets and display the main ones as a function of the observing epoch in Fig. 3. Notably, we find that the large-scale field progressively weakens from 2019B to 2022B before increasing until the end of the campaign. The inferred magnetic topology, almost fully poloidal and axisymmetric at all epochs, features a dominant 1.1–1.4 kG dipole component inclined at 10–20° to the rotation axis. This dipolar component contains ≃70% of the reconstructed magnetic energy at all epochs and varies both in strength (minimum in 2022B) and orientation with time. The average small-scale field, <B_s>, that we derive from these reconstructions (from a weighted average over the visible stellar hemisphere, taking into account limb-darkening; see column 4 of Table C.1) predicts a slightly larger epoch-to-epoch variation. Otherwise, it is consistent with observations (decreasing from 2019 to 2022–2023 before increasing in 2024; see Fig. 1) and with the full-amplitude modulation at each epoch (minimum in 2020 and maximum in 2022). In addition, ZDI suggests that the largest Zeeman broadening arises from the strongest field near the poles, similar to previous results from Stokes I data on another young active star (Kochukhov et al. 2023).

We note that attempting to simultaneously fit all data collected in a single season (e.g., 2024A and 2024B) always yields a larger ${\chi }_{\mathrm{r}}^2$ and thus a significantly worse fit than that obtained when splitting each season in half, as temporal evolution (including differential rotation) becomes quite notable on a timescale of up to 207 d (in 2024). In fact, Stokes V LSD profiles within the half-season subsets already show evidence for temporal variability beyond rotational modulation, with slightly discrepant Zeeman signatures detected at almost equal rotational phases but different cycles (e.g., 12.336 and 20.354 in 2024; see bottom right panel of Fig. C.1). This is actually expected given the evolution timescale θ₃ derived from the B_ℓ analysis of Sect. 3 (101 ± 9 d, see Table B.1), which is comparable to the maximum length of the subsets but much shorter than the duration of the observing seasons. We also note that the field strengths reconstructed in this new study are approximately twice as large as those presented in the previous study (Donati et al. 2023a) at the same epochs, mostly reflecting the much tighter fit to the Stokes I profiles achieved in this new work. As a result, the new inferred magnetic topologies yield small-scale field values (through a weighted average of the reconstructed field strengths over the visible hemisphere; see above) that are much more consistent with actual measurements and agree well with the results of the latest magnetometric study of AU Mic from Stokes IV QU data collected in 2023B (Donati et al. 2025a).

Fig. 1 Longitudinal magnetic field Bℓ (top panel), small-scale magnetic field <B> (middle panel) and temperature variations dT (bottom panel) of AU Mic (red dots), with the QP GPR fit to the data (solid cyan line) and corresponding 68% confidence intervals (dotted cyan lines). The residuals, shown in the lower plot of each panel, yield rms values of 7.2 G, 0.038 kG and 0.96 K (${\chi }_{\mathrm{r}}^2$ = 2.3, 0.98 and 0.60, respectively). A zoom on the 2023 and 2024 data is shown in Figs. B.1 and B.2.

5 Radial velocities of AU Mic

In this section, we simultaneously analyze the full set of RVs derived for AU Mic from 2019 to 2024, repeating most of the previous analysis on the complete set of 344 measurements, which now spans 2041 d in the life of the AU Mic multi-planet system. As mentioned in Sect. 2, the new RVs were derived with the latest versions of APERO and LBL, which implement a more accurate correction of telluric contamination and are expected to yield even more precise RV estimates.

Following Donati et al. (2023a), we analyzed three main cases: a reference scenario including transiting planets b and c; a second case including b, c and the candidate outer planet e (the existence of which was recently challenged by Mallorquín et al. 2024); and a third case including b, c, e, and the putative planet d, located between b and c and presumably causing the reported transit time variations (TTVs) of planets b and c (Wittrock et al. 2023; Boldog et al. 2025). For each of these three cases, we fit the observed RV curve with a model including a dominant activity signal from the star, described with a QP GP (as in Sect. 3), as well as a Keplerian signal for the considered planets. Although we know that transiting planets b and c indeed exist, we nonetheless included a fourth case in which no planets were included, to determine how well their RV signatures are detected and to compare with the previous results. In each case, the free model parameters and corresponding error bars were inferred through an MCMC process (as in Sect. 3).

Since orbital periods and transit (or conjunction) times for b, c, and d have already been accurately determined from photometric and TTV data (e.g., Szabó et al. 2022; Wittrock et al. 2023; Mallorquín et al. 2024), we chose to fix them, which is virtually equivalent to allowing them to vary within the very narrow prior determined by previous studies. Assuming circular orbits for all planets (as a first step), we are left with six planet parameters: the RV semi-amplitudes of all four planets (denoted K_b, K_c, K_d, and K_e), and the orbital period and conjunction time of planet e (P_e and BJD_e). Of these six planet parameters, we respectively fit zero, two (K_b, K_c), five (all but K_d) or all six from the RV data in the four considered cases, in addition to the five GP hyper parameters describing the activity. The values and error bars derived from the MCMC process for the relevant parameters are listed in Table 1, along with the corresponding priors, while the best fit to the observed RVs is shown in Fig. 4 (with a zoom on the 2023 and 2024 data in Fig. D.1), for the model including all four planets. We find that the model without planets yields a significantly worse fit to the SPIRou RVs of AU Mic than that including planets b and c (Δ log ℒ_M = −12.4). In the other cases, the average semi-amplitudes derived for planets b and c are K_b = ${2.8}_{-0.8}^{+1.1}$ m s⁻¹ and K_c = ${3.9}_{-0.9}^{+1.1}$ m s⁻¹, and correspond to masses of M_b = ${6.3}_{-1.8}^{+2.5}$ M_⊕ and M_c = ${11.6}_{-2.7}^{+3.3}$ M_⊕ (where M_⊕ is the Earth’s mass) and to detection levels of 3.5 and 4.3σ, respectively. These semi-amplitudes are smaller than, but still consistent within error bars with, those inferred from the previous study, as well as with the recent estimates from the joint analysis of CARMENES, HARPS and the first set of SPIRou data (K_b = 3.6 ± 1.1 m s⁻¹ and K_c = 4.3 ± 1.0 m s⁻¹, Mallorquín et al. 2024).

We also find that including candidate planet e significantly improves the fit (Δ log ℒ_M = 11.5) although the derived semi-amplitude, equal to K_e = ${5.9}_{-1.2}^{+1.5}$ m s⁻¹, is approximately half as large as that derived in Donati et al. (2023a), but still compatible within ≃2.5σ. This would imply that candidate planet e, if real, has a mass of M_e = ${21.1}_{-4.3}^{+5.4}$ M_⊕ and is detected at a level of 4.9σ. At P_e = 33.11 ± 0.06 d, its orbital period is also slightly shorter (by about 0.3 d or 2.5σ) than the previous estimate, but remains consistent with orbits that would potentially be stable on a Gyr timescale (for small eccentricities, see Fig. 12 of Donati et al. 2023a), i.e., much longer than the age of the AU Mic system, assuming all planets are coplanar. Figure 5 shows the periodogram of the raw, activity-filtered, and residual RVs, with the peak corresponding to candidate planet e associated with a false alarm probability (FAP) that the signal is spurious of about 2 × 10⁻⁹ in the filtered RVs. Figure 6 presents a stacked periodogram of the filtered RVs, showing a signal at P_e that consistently strengthens as more data are added (rather than varying up and down like a spurious signal from activity), suggesting that candidate planet e is real. The updated value of K_e is also significantly lower than the 3.5σ upper limit of 10 m s⁻¹ derived by Mallorquín et al. (2024), potentially explaining their non-detection. Figure 7 shows the phase-folded RV curves of b, c, and e. Including candidate planet d only marginally improves the likelihood (Δ log ℒ_M = 0.3) with respect to the b+c+e case, without significantly changing the results for the other three planets. The 90 and 99% confidence upper limits on K_d are 1.9 and 2.7 m s⁻¹ (4.9 and 7.0 M_⊕ for M_d, respectively). Running the same experiment with Keplerian (i.e., non-circular) orbits for transiting planets b and c, and for candidate planet e, yields only a marginal improvement (Δ log ℒ_M ≃1.5), confirming a posteriori that circular orbits are the most likely scenario for all three planets, with respective error bars on the eccentricities equal to 0.02, 0.12, and 0.15 for b, c, and e.

Finally, we note that the hyperparameters of the GP describing the activity of AU Mic over the full campaign are similar in all four cases. The average semi-amplitude θ₁ is ≃40 m s⁻¹, i.e., 25% larger than in the previous study, although still a factor of 2.5 to 4 smaller than that at optical and visible wavelengths (Mallorquín et al. 2024). At ≃167 d, the evolution timescale, θ₃, is also about 20% longer than the earlier estimate and is fully consistent with that derived from dT (see the bottom section of Table B.1). It is also much longer than the orbital periods of b, c, d, and e, suggesting minimal interference between the GPR fit and the planet parameters. The recurrence period, θ₂ (another proxy for P_rot), is marginally longer than that obtained from B_ℓ and dT, but consistent with that derived from <B>. All four measurements are associated with very small error bars, ranging from 0.0014 to 0.0019 d (2.0 to 2.7 min). We also report slight seasonal changes in the GP parameters, similar to those found for B_ℓ. In particular, in 2020, the evolution timescale (θ₃ = 129 ± 40 d) and recurrence period (θ₂ = 4.850 ± 0.007 d) are marginally shorter than in the other seasons.

Fig. 2 Reconstructed maps of the large-scale field of AU Mic showing the radial, azimuthal and meridional components in spherical coordinates (left, middle and right columns, units in G), for season 2023A, 2023B, 2024A and 2024b (top to bottom rows, respectively). These maps, derived from the Stokes IV LSD profiles of Fig. C.1 using ZDI, are displayed in a flattened polar projection down to latitude −60°, with the north pole at the center and the equator depicted as a bold line. Outer ticks mark the phases of observations. Positive radial, azimuthal, and meridional fields point outwards, counterclockwise, and polewards, respectively.

Fig. 3 Quadratic average of the large-scale magnetic field over the stellar surface (red; column 2 of Table C.1), polar strength of the dipolar component (green; column 5 of Table C.1), and average small-scale field over the rotation cycle (blue; column 4 of Table C.1) as a function of the observing epoch, for the 11 magnetic topologies of AU Mic derived with ZDI.

Fig. 4 Raw (top), filtered (middle), and residual (bottom) RVs of AU Mic (red dots) over the observing period. The top panel shows the MCMC fit to the data, including a QP GPR modeling of the activity and the RV signatures of all four planets (cyan). The middle panel shows the planet RV signatures (pink, blue, green, orange, and cyan for planets b, c, d, e, and b+c+d+e, respectively) once activity is filtered out. The rms of the residuals is 10.1 m s−1. A zoom on the 2023 and 2024 data is shown in Fig. D.1.

Fig. 5 Periodogram of the raw (top), filtered (middle), and residual (bottom) RV data, including all planets in the MCMC modeling. Dashed vertical cyan lines trace the rotation period of the star and the planet orbital periods; the dashed horizontal line indicates the 0.1% FAP level in the periodogram of the RV data. The peak corresponding to candidate planet e (with a 1-yr alias at 30.3 d) dominates the middle panel, with a FAP of 2 × 10−9. The orange curve shows the window function, which peaks at the synodic period of the Moon (at 29.5 d).

6 Activity of AU Mic

We also investigated the temporal behavior of several additional activity proxies (beyond those already mentioned in Sect. 3), which we briefly discuss below.

In addition to providing precise RVs (through the coefficient of the first term in the Taylor expansion used to describe the variational profile of each line with respect to its median; see Artigau et al. 2022), LBL also yields estimates of the changes in line width and asymmetry (through the coefficients of the second and third terms in the same Taylor expansion and denoted d2v and d3v). These quantities share similarities with the more conventional proxies known as differential line width and bisector span (dLW and BIS), computed from the cross-correlation profile of spectral lines. Fitting d2v and d3v with GPR (as for B_ℓ and dT in Sect. 3), we find that both quantities are rotationally modulated, with recurrence periods consistent within error bars with that derived from RVs. We also find that dT correlates reasonably well with d2v (R=0.66), though less tightly than with <B> (see Sect. 3) and that RVs are anticorrelated with d3v (R=−0.60), as is usually the case with BIS when RVs are dominated by activity jitter. The strongest correlation with RVs is, however, not with d2v or d3v (or their first time derivatives), but with the first time derivative of dT (R=−0.85). Even then, exploiting this correlation to filter RVs from the activity jitter still leaves a residual activity signal with a semi-amplitude of about 40% that of the original activity signal (similar to the results of the first study, see Donati et al. 2023a).

We also attempted a multidimensional GPR fit to our RV and dT data simultaneously (as advocated by Rajpaul et al. 2015) but still ended up with a residual RV activity signal of semiamplitude ≃1/3 of the original, likely reflecting the limitation of the F F′ approach underlying the multidimensional GPR modeling (see Tables 1 and D.1). We note that the multidimensional modeling significantly reduced the GP evolution timescale (from ≃170 to ≃120 d) in an attempt to compensate for this limitation, although with moderate success. The resulting fit to the RVs is significantly worse than with a simple GP, with a 2.4× larger ${\chi }_{\mathrm{r}}^2$, a 1.55× higher rms, and a 1.4× higher excess RV jitter. Despite this limitation, the parameters inferred for planets b, c, and candidate planet e, remain consistent with those derived in Sect. 5 within <1σ (albeit with larger error bars). The corresponding log BF values with respect to the two-planet reference case are also consistent (see Table D.1). We thus conclude that, for stars with dominant RV activity signal like AU Mic, straightforward GPR analyses of RVs remain the most efficient approach to filter densely sampled RV curves from the activity jitter.

We also examined the 1083 nm He I and 1282 nm Paβ lines, known as reliable chromospheric activity proxies in low-mass stars. To investigate this, we followed Donati et al. (2023a), i.e., normalizing each He I and Paβ profile by the corresponding median of all recorded profiles, then fitting the normalized profile by a Gaussian of fixed full width at half maximum (FWHM=40 km s⁻¹) centered on the stellar rest frame, to find out how much each line changed with respect to the median at all epochs. Despite AU Mic being quite active, we find that both lines are only weakly variable, with equivalent width fluctuations of ≃0.8 km s⁻¹ rms for He I and ≃0.3 km s⁻¹ rms for Paβ over the full campaign. Most of this dispersion is attributable to intrinsic variability, with a few major eruptive events occurring from time to time (e.g., on 2019 June 14, 2020 May 30, 2021 November 13, 2023 August 28, and 2024 September 22), during which both lines can appear in emission, with broad wings extending beyond ±100 km s⁻¹ in the stellar rest frame and equivalent widths exceeding 15 km s⁻¹. Outside of flares, rotational modulation is weak, consistent with the largely axisymmetric magnetic topology, with average semi-amplitudes in equivalent width variations of about 0.2 and 0.1 km s⁻¹ for He I and Paβ, respectively. Nonetheless, the modulation is strong enough to dominate the periodogram of both lines over the full range of the observations. We also verified that flares, even major ones, do not have a clear impact on the measured RVs (at an rms level of ≃10 m s⁻¹), with residuals to the GPR fit that are consistent with the bulk of the data points. A GPR fit excluding RVs at flaring epochs yields results consistent within 0.5σ with those including all RV points.

Focusing on the fluctuations of the He I line, which are more clearly detected than those of Paβ, and computing a 2D periodogram for the 11 subsets outlined in Sect. 4, we find a dominant peak at P_rot in subset 2019B only (see Fig. E.1, left panel) but in no other subsets. Otherwise P_rot is at best weak (as in, e.g., 2023B) and most of the time undetected, even when spectra featuring large flares or those most strongly contaminated with telluric lines are excluded, reflecting the high level of intrinsic variability relative to rotational modulation. In 2019 B, maximum He I emission occurs at phase 0.92, i.e., slightly before the positive magnetic pole is furthest from the observer (at phase 0.02, see Table C.1), as expected from the magnetic equator being brighter than the pole at the chromospheric level. We also note one subset (2020B) where significant power is detected at the orbital period of planet b (see Fig. E.1, right panel), potentially suggesting energetic star-planet interactions at this particular epoch with maximum He I emission occurring 2.4 d after the transit. However, we stress that this detection, although present across the whole He I triplet as expected from a real signal, may actually be coincidental, as similar power levels are also seen at larger periods and outside the stellar line at this epoch (as well as in others). The high degree of intrinsic variability in the He I line of AU Mic, coupled to the uneven sampling of the observations, may indeed generate spurious signals appearing at random periods in some subsets.

Fig. 6 Stacked periodograms of the filtered RVs, as a function of the number of RV points included in the Fourier analysis, beginning from the first observation. The color scale indicates the logarithmic power in the periodogram. The main RV signal associated with candidate planet e, outlined with a vertical dashed line (see Table 1), becomes stronger and increasingly dominant in this period range as more spectra are added to the analysis. The horizontal dashed line illustrates the end of the previous data set (Donati et al. 2023a). The weaker peak at 30.3 d, also visible in the middle panel of Fig. 5, is a 1-yr alias of the main signal.

Fig. 7 Phase-folded filtered (top plots) and residual (bottom plots) RVs for transiting planets b (top panel) and c (middle panel), and for candidate planet e (bottom panel) of AU Mic. Red dots show the individual RV measurements with their error bars, while black stars indicate average RVs over 0.1 phase bins. As in Fig. 4, the dispersion of RV residuals is 10.1 m s−1.

7 Summary and discussion

In this paper, we analyzed an extended data set of 382 unpolarized and circularly-polarized spectra of AU Mic collected with SPIRou over a total timespan of 2041 d from early 2019 to late 2024. This set included 157 new observations recorded in late 2022, 2023, and 2024, which add to the 225 already outlined and analyzed in the initial study (Donati et al. 2023a). We carried out a similar analysis of the large-scale and small-scale magnetic field of AU Mic, its activity and its multi-planet system, reassessing in particular the masses of the two transiting planets (b and c), and the potential existence of the candidate outer planet (e), which was proposed by Donati et al. (2023a) but challenged by Mallorquín et al. (2024). From these SPIRou spectra, we derived time series of: LSD Stokes I and V profiles; B_ℓ values (computed from LSD signatures); <B> estimates (inferred from the broadening of magnetically sensitive lines in SPIRou spectra with ZeeTurbo); and RVs and dT measurements (both computed with LBL, see Sect. 2).

We first find that the evolution of B_ℓ and <B> is similar to that previously described, with a rotation period now refined to a precision better than 0.002 d, a factor of about two better than the uncertainty in our previous estimate. We find that B_ℓ exhibited another minimum in the rotational modulation amplitude in 2023, but has not yet regained the maximum amplitude observed in 2020, nor has it changed its average sign (B_ℓ being, on average, positive over all seasons to date). Similarly, <B> steadily decreased (from 2.81 to 2.64 kG) over most of the observing period and only began to increase during the final season. However, the rotational modulation has not yet reached the low amplitude observed in 2020. In addition, we confirm the previous finding that neither B_ℓ nor |B_ℓ| correlate with <B>. In contrast, <B> is strongly anticorrelated with dT, dT decreasing at a rate of 76 K/kG as <B> increases (magnetic regions being cooler than the quiet photosphere). These results suggest that, if unspotted, AU Mic would exhibit a photospheric temperature of ≃3875 K. They also indicate that the average spot coverage at the stellar surface was about 34% at the time of monitoring (assuming a typical spot-to-photosphere temperature contrast of 620 K for AU Mic, Berdyugina 2005), with rotational modulation and seasonal changes inducing periodic fluctuations of ±1–2%.

By analyzing the rotational modulation of the Stokes IV LSD profiles of AU Mic with ZDI over the 11 subsets we defined (each covering up to 120 d), we find that the large-scale topology of AU Mic is mainly poloidal with a quadratic average of ≃1 kG. It primarily consists of a 1.1–1.4 kG dipole tilted at 10–20° to the rotation axis (see Table C.1), which contains 70% of the reconstructed magnetic energy. The small-scale surface field, <B_I>, derived from our self-consistent ZDI modeling and based on the assumption that <B_I> locally scales up with the large-scale field <B_V> at a rate of f_I/f_V = 4.5 (see Sect. 4), reaches up to 5 kG at the pole. Its average over the visible hemisphere <B_s> varies from 2.5 to 3.2 kG, in agreement with the <B> estimates with ZeeTurbo, both in terms of long-term evolution and rotational modulation amplitude (see Sects. 3 and 4). The inferred dipole strength is about twice that found in the previous study, reflecting a tighter ZDI fit to the Stokes I LSD profiles. The dipole intensity varies similarly to <B>, i.e., decreasing from 2019 to a minimum in late 2022, then increasing toward the end of the observations. These magnetic properties of AU Mic and their variations with time are consistent with those reported for magnetic fields of active M dwarfs in previous studies (Morin et al. 2008; Kochukhov 2021; Reiners et al. 2022).

If the magnetic field of AU Mic follows a Sun-like periodic cycle, its duration must be much longer than our six-yr monitoring. This suggests that the five-yr activity cycle proposed by Ibañez Bustos et al. (2019) may represent only a transient feature in a longer evolution. The field evolution of AU Mic is clearly distinct from, and more complex than, that of less active M dwarfs for which evidence of Sun-like cycles, including global polarity reversals of their poloidal field, has been reported (Donati et al. 2023b; Lehmann et al. 2024). The magnetic fluctuations of AU Mic more closely resemble those of the other young (more massive) planet-hosting star V1298 Tau, whose large-scale field was shown to evolve even more rapidly, showing neither periodicity nor polarity switches, at least over timescales of several years (Finociety et al. 2023). Similarly, other reports based on spectropolarimetric data spanning up to 15 yr (e.g., Bellotti et al. 2024) similarly suggest that rapidly rotating M dwarfs host large-scale fields that, if periodic, have cycles longer than a decade, or are intrinsically non periodic.

AU Mic and its large-scale magnetic topology are conveniently oriented for Earth-based observers to detect and monitor radio emission from the stellar magnetosphere, particularly the reported highly circularly polarized and rotationally modulated bursts of coherent emission. These bursts are thought to be generated by the electron cyclotron maser instability (Bloot et al. 2024), likely occurring in the auroral rings of the magnetic poles (Hallinan et al. 2015). The dipole field strength we infer is consistent with estimates derived from the frequencies at which radio bursts are observed (yielding 0.39–1.1 kG, Bloot et al. 2024), which implies an emission region located up to ≃0.5 R_⋆ above the surface if generated at the fundamental cyclotron frequency, or up to ≃0.9 R_⋆ if produced at the first harmonic.

The activity of AU Mic, estimated from the He I and Paβ proxies, features a dominant level of stochastic variability on top of a weak amount of rotational modulation, and 2D periodograms of these spectral lines in the individual subsets of the observations rarely show a clear signal at the rotation period.

We also used the SPIRou data to further constrain the masses of the planets in the AU Mic system, including the two transiting planets b and c, the candidate planet d (proposed to explain the reported TTVs of b and c, Wittrock et al. 2023; Boldog et al. 2025) and the outer candidate planet e, whose existence was previously inferred (Donati et al. 2023a). For planets b and c, we derived RV semi-amplitudes of K_b = ${2.8}_{-0.8}^{+1.1}$ m s⁻¹ and K_c = ${3.9}_{-0.9}^{+1.1}$ m s⁻¹, respectively, corresponding to detection levels of 3.5 and 4.3σ and to masses of M_b = ${6.3}_{-1.8}^{+2.5}$ M_⊕ and M_c = ${11.6}_{-2.7}^{+3.3}$ M_⊕. These values are lower but remain consistent with the most recent estimates from Mallorquín et al. (2024). We find no evidence that either planet has a non-circular orbit, with respective error bars on their eccentricities equal to 0.02 and 0.12 for b and c. While the orbit of planet b was shown to be well aligned with the equatorial plane of the star (Palle et al. 2020), planet c may be on a misaligned orbit (Yu et al. 2025). Such a scenario would require c, presumably on an aligned orbit when still migrating within the inner disc, to have switched to another orbit at some point while maintaining circularity (e.g., through a resonance with an outer planet) and dynamical stability (unlikely for severe misalignements; Yu et al. 2025).

Using recent radius estimates for the two planets (R_b = 4.79 ± 0.29 R_⊕ and R_c = 2.79 ± 0.18 R_⊕ where R_⊕ is the Earth’s radius; Mallorquín et al. 2024, consistent with the latest measurements from Boldog et al. 2025), we find respective densities of ${0.32}_{-0.10}^{+0.13}$ and ${2.9}_{-0.8}^{+1.1}$ g cm⁻³, i.e., a factor of ≃10 higher for c than for b. While planet c falls in the middle of the main exoplanet sequence in a radius versus mass diagram (with densities in the range 1–10 g cm⁻³), b lies well above this sequence (see Fig. 8), suggesting that it has not completed its contraction yet. This could result from its higher equilibrium temperature (possibly boosted by induction heating; Kislyakova et al. 2018) and its lower mass. Alternatively, planet b may have a different composition than c, as reported for other systems (e.g., Barat et al. 2024), possibly reflecting a different formation process, if the orbit of c is indeed misaligned. In terms of density, planet b is similar to the youngest known close-in planet, recently discovered around the classical T Tauri star IRAS 04125+2902 (Barber et al. 2024), whose bulk density is even lower (<0.23 g cm⁻³; Donati et al. 2025b). Being inflated and expected to lose mass at a rate 10× larger than that of planet c (Mallorquín et al. 2024), b is especially interesting for characterizing the structure, chemistry, and dynamics of young planet atmospheres (Barat et al. 2024). However, attempts to detect and estimate atmospheric escape from planet b through transit spectroscopy have so far only yielded upper limits in the 1083 nm He I triplet (Hirano et al. 2020; Allart et al. 2023; Masson et al. 2024), or a variable detection in Ly α (Rockcliffe et al. 2023). Stochastic activity from the host star limits the precision with which transit signatures can be characterized.

In addition, we confirm that the RV signature of candidate planet e, discussed in the previous study, remains present in the extended data set with a confidence rate high enough to qualify as a reliable detection (Δ log ℒ_M = 11.5), albeit with a lower semi-amplitude of K_e = ${5.9}_{-1.2}^{+1.5}$ m s⁻¹ yielding a mass of ${21.1}_{-4.3}^{+5.4}$ M_⊕. The derived orbital period, P_e = 33.11 ± 0.06 d, corresponds to a distance of 0.17 au from the host star, with the signal at P_e emerging clearly in the 2D stacked periodogram of the filtered RVs as more data points are progressively included in the analysis. The RV data show no evidence that the orbit of planet e departs from being circular (with an error bar of 0.15 on the eccentricity), a condition likely required for dynamical stability over Gyr timescales (Donati et al. 2023a), assuming all planets are coplanar. Since no transit of planet e has been reported in the literature, we suspect that it does not transit. This implies that the axis of its orbital plane is inclined by no more than 88.7° with respect to the line of sight, a value consistent with the orbital plane inclination of planets b and c (Mallorquín et al. 2024). If planet c is misaligned (Yu et al. 2025), it may have become so under the influence of planet e when both entered a 2:1 resonance in the late phases of their migration process. If candidate planet d, proposed to explain the TTVs of b and c, is indeed present in the system, we derive 90 and 99% confidence upper limits on M_d of 4.9 and 7.0 M_⊕ respectively (assuming P_d = 12.73596 d and a transit BJD of 2458340.55781). These limits are consistent with the estimate of Wittrock et al. (2023, 1.0 ± 0.5 M_⊕), or the much lower one of Boldog et al. (2025, 0.1 M_⊕).

Given the intensity of the large-scale field, particularly its dipole component, the Alfven volume (in which the magnetic pressure dominates over the dynamic pressure) associated with AU Mic’s wind likely extends farther than initially estimated (Kavanagh et al. 2021), by up to 30% assuming a similar topology, though possibly less for a more axisymmetric configuration (Alvarado-Gómez et al. 2022). Following Kavanagh et al. (2021), we find that it may thus include the orbit of planet b for a stellar wind with a mass-loss rate up to 1000 ${\dot{\mathrm{M}}}_{\odot }$ (where ${\dot{\mathrm{M}}}_{\odot }$ is the mass loss rate of the Sun, i.e., 2 × 10⁻¹⁴ M_⊙ yr⁻¹) and that of c for a mass-loss rate up to ≃250 ${\dot{\mathrm{M}}}_{\odot }$. In this context, star-planet interactions may occur between AU Mic and its innermost planet b, and likely also with planets c and d, possibly even with e if the mass-loss rate is ≤100 ${\dot{\mathrm{M}}}_{\odot }$. This may explain the activity signatures observed at times in the 1083 nm He I triplet at the orbital period of b (see right panel of Fig. E.1), similar to previous claims involving the 587 nm He I line in the visible spectrum of AU Mic at the same epoch (Klein et al. 2022). It is unclear, however, why the maximum He I flux would occur as much as 2.4 d after transit, or why the He I periodogram should peak at the orbital period rather than at the synodic period (of 11.4 d), if what we see corresponds to emission at the stellar surface from the footpoints of field lines connecting the star to planet b. Moreover, if the reported He I emission were indeed caused by star-planet interactions, it would be expected to persist over time, given the relatively stable and axisymmetric large-scale magnetic configuration of AU Mic. We therefore conclude that the modulated activity signal detected in the He I line at the orbital period of planet b is more likely to reflect the combined effect of intrinsic variability and irregular sampling.

Given its youth, its multi-planet system and its proximity to the Earth, AU Mic is an ideal target for spectroscopic, spectropolarimetric, and photometric observations on a long-term monitoring basis. Such observations will help to further document the orbital and atmospheric properties of its close-in planets, the magnetic field and activity of the host star, potential interactions taking place between them, and the impact of massive coronal mass ejections from the host stars on the planets (Alvarado-Gómez et al. 2022). We therefore advocate the stellar and exoplanet communities to continue dedicating observing time to the study of this key object from both ground-based and spaceborne facilities in the coming years.

Fig. 8 Mass-radius diagram for exoplanets with both mass and radius known with a relative precision better than 30% (gray points). AU Mic b and c are shown with red circles and error bars. Theoretical models from Zeng et al. (2019) for various inner planet structures and compositions are depicted with a solid black line (100% iron and Earth-like) and blue line (50% water envelope). Models from Aguichine et al. (2021) for a 50% water envelope including the effects of irradiation, are plotted as a solid green line. Models with a 1,2 and 5% H2 atmosphere with either an Earth-like (dashed brown lines) or a 50% water (orange dashed line) interior are also shown. Dashed grey lines indicate iso-densities of 1,3, and 10 g cm−3 (from top to bottom).

Data availability

Data used in this study were collected within the SLS (up to mid-2022) and the SPICE LPs, and additional PI and DDT programs. The SLS, DDT and pre-2023 PI data are publicly available from the Canadian Astronomy Data Center (https://www.cadc-ccda.hia-iha.nrc-cnrc.gc.ca). Data from the SPICE LP and the 2024 PI program will be available from 2025 September and 2026 March, respectively. Full Tables A.1 and A.2 are available at the CDS via https://cdsarc.cds.unistra.fr/viz-bin/cat/J/A+A/700/A227.

Acknowledgements

We thank an anonymous referee for valuable comments that improved the manuscript. This study is based on data obtained at the CFHT, operated by the CNRC (Canada), INSU/CNRS (France) and the University of Hawaii. The authors wish to recognise and acknowledge the very significant cultural role and reverence that the summit of Maunakea has always had within the indigenous Hawaiian community. This work also benefited from the SIMBAD CDS database at http://simbad.u-strasbg.fr/simbad and the ADS system at https://ui.adsabs.harvard.edu.

Appendix A Observation log

Tables A.1 and A.2 provide the observation log for the Libre ESpRIT and APERO spectra respectively, and the measurements derived from them at each epoch.

Appendix B Magnetic field and temperature variations of AU Mic: additional material

Figures B.1 and B.2 show the B_ℓ, <B> and dT curves and corresponding GPR fits, respectively zooming on the 2023 and 2024 data, whereas Table B.1 details the results of the GPR fit to the overall B_ℓ, <B> and dT data.

Fig. B.1 Same as Fig. 1, zooming on the 2023 data.

Fig. B.2 Same as Fig. 1, zooming on the 2024 data.

Appendix C ZDI modeling of AU Mic: additional material

Figure C.1 shows the observed and reconstructed LSD Stokes I and V profiles of AU Mic for seasons 2023A to 2024B, whereas Figs. C.2 and C.3 show the reconstructed ZDI maps of AU Mic for seasons 2019B to 2022B. Table C.1 recaps the main characteristics of the magnetic topologies reconstructed with ZDI for all subsets.

Fig. C.1 Observed (thick black line) and modelled (thin red line) LSD Stokes I (top panels) and V (bottom panels) profiles of the photospheric lines of AU Mic, for subsets 2023A, 2023B, 2024A and 2024B (from left to right). The ZDI modeling of these profiles is described in Sec. 4. Rotation cycles (counting from 219, 244, 295 and 316 for 2023A, 2023B, 2024A and 2024B respectively, see Table A.1) are indicated to the right of all profiles, whereas ±1σ error bars are shown to the left of LSD Stokes V profiles.

Fig. C.2 Same as Fig. 2 for seasons 2019B, 2020A, 2020B, 2021A and 2021B.

Fig. C.3 Same as Fig. 2 for seasons 2022A and 2022B.

Appendix D RVs of AU Mic: additional material

Figure D.1 shows the RV curve of AU Mic and the corresponding 4-planet and GPR fits, zooming on the 2023 and 2024 data.

Fig. D.1 Same as Fig. 4, zooming on the 2023 (top) and 2024 (bottom) data.

Table D.1 details the results of the multi-dimensional GPR fit to the dT and RV data mentioned in Sec. 6.

Appendix E Activity of AU Mic: additional material

Figure E.1 depicts the 2D periodograms of the 1083.3 nm He I line in the spectrum of AU Mic, in semesters 2019B and 2020B.

Fig. E.1 2D periodograms of the 1083.3-nm He I triplet residual (with respect to the median) in the stellar rest frame, for the 2019B (left) and 2020B (right) spectra of AU Mic. In both periodograms, the dashed horizontal line traces Prot and the orbital periods of the transiting and candidate planets, whereas the vertical dotted lines depict the velocities of the three components of the He I triplet. The color scale traces the logarithmic power in the periodograms. We caution that only the main peaks (colored yellow to red and extending over at least several velocity bins) are likely to be significant in these plots.

References

All Tables

All Figures

Fig. 1 Longitudinal magnetic field Bℓ (top panel), small-scale magnetic field <B> (middle panel) and temperature variations dT (bottom panel) of AU Mic (red dots), with the QP GPR fit to the data (solid cyan line) and corresponding 68% confidence intervals (dotted cyan lines). The residuals, shown in the lower plot of each panel, yield rms values of 7.2 G, 0.038 kG and 0.96 K (${\chi }_{\mathrm{r}}^2$ = 2.3, 0.98 and 0.60, respectively). A zoom on the 2023 and 2024 data is shown in Figs. B.1 and B.2.

In the text

Fig. 2 Reconstructed maps of the large-scale field of AU Mic showing the radial, azimuthal and meridional components in spherical coordinates (left, middle and right columns, units in G), for season 2023A, 2023B, 2024A and 2024b (top to bottom rows, respectively). These maps, derived from the Stokes IV LSD profiles of Fig. C.1 using ZDI, are displayed in a flattened polar projection down to latitude −60°, with the north pole at the center and the equator depicted as a bold line. Outer ticks mark the phases of observations. Positive radial, azimuthal, and meridional fields point outwards, counterclockwise, and polewards, respectively.

In the text

	Fig. 3 Quadratic average of the large-scale magnetic field over the stellar surface (red; column 2 of Table C.1), polar strength of the dipolar component (green; column 5 of Table C.1), and average small-scale field over the rotation cycle (blue; column 4 of Table C.1) as a function of the observing epoch, for the 11 magnetic topologies of AU Mic derived with ZDI.
In the text

Fig. 4 Raw (top), filtered (middle), and residual (bottom) RVs of AU Mic (red dots) over the observing period. The top panel shows the MCMC fit to the data, including a QP GPR modeling of the activity and the RV signatures of all four planets (cyan). The middle panel shows the planet RV signatures (pink, blue, green, orange, and cyan for planets b, c, d, e, and b+c+d+e, respectively) once activity is filtered out. The rms of the residuals is 10.1 m s−1. A zoom on the 2023 and 2024 data is shown in Fig. D.1.

In the text

Fig. 5 Periodogram of the raw (top), filtered (middle), and residual (bottom) RV data, including all planets in the MCMC modeling. Dashed vertical cyan lines trace the rotation period of the star and the planet orbital periods; the dashed horizontal line indicates the 0.1% FAP level in the periodogram of the RV data. The peak corresponding to candidate planet e (with a 1-yr alias at 30.3 d) dominates the middle panel, with a FAP of 2 × 10−9. The orange curve shows the window function, which peaks at the synodic period of the Moon (at 29.5 d).

In the text

Fig. 6 Stacked periodograms of the filtered RVs, as a function of the number of RV points included in the Fourier analysis, beginning from the first observation. The color scale indicates the logarithmic power in the periodogram. The main RV signal associated with candidate planet e, outlined with a vertical dashed line (see Table 1), becomes stronger and increasingly dominant in this period range as more spectra are added to the analysis. The horizontal dashed line illustrates the end of the previous data set (Donati et al. 2023a). The weaker peak at 30.3 d, also visible in the middle panel of Fig. 5, is a 1-yr alias of the main signal.

In the text

	Fig. 7 Phase-folded filtered (top plots) and residual (bottom plots) RVs for transiting planets b (top panel) and c (middle panel), and for candidate planet e (bottom panel) of AU Mic. Red dots show the individual RV measurements with their error bars, while black stars indicate average RVs over 0.1 phase bins. As in Fig. 4, the dispersion of RV residuals is 10.1 m s−1.
In the text

Fig. 8 Mass-radius diagram for exoplanets with both mass and radius known with a relative precision better than 30% (gray points). AU Mic b and c are shown with red circles and error bars. Theoretical models from Zeng et al. (2019) for various inner planet structures and compositions are depicted with a solid black line (100% iron and Earth-like) and blue line (50% water envelope). Models from Aguichine et al. (2021) for a 50% water envelope including the effects of irradiation, are plotted as a solid green line. Models with a 1,2 and 5% H2 atmosphere with either an Earth-like (dashed brown lines) or a 50% water (orange dashed line) interior are also shown. Dashed grey lines indicate iso-densities of 1,3, and 10 g cm−3 (from top to bottom).

In the text

	Fig. B.1 Same as Fig. 1, zooming on the 2023 data.
In the text

	Fig. B.2 Same as Fig. 1, zooming on the 2024 data.
In the text

Fig. C.1 Observed (thick black line) and modelled (thin red line) LSD Stokes I (top panels) and V (bottom panels) profiles of the photospheric lines of AU Mic, for subsets 2023A, 2023B, 2024A and 2024B (from left to right). The ZDI modeling of these profiles is described in Sec. 4. Rotation cycles (counting from 219, 244, 295 and 316 for 2023A, 2023B, 2024A and 2024B respectively, see Table A.1) are indicated to the right of all profiles, whereas ±1σ error bars are shown to the left of LSD Stokes V profiles.

In the text

	Fig. C.2 Same as Fig. 2 for seasons 2019B, 2020A, 2020B, 2021A and 2021B.
In the text

	Fig. C.3 Same as Fig. 2 for seasons 2022A and 2022B.
In the text

	Fig. D.1 Same as Fig. 4, zooming on the 2023 (top) and 2024 (bottom) data.
In the text

Fig. E.1 2D periodograms of the 1083.3-nm He I triplet residual (with respect to the median) in the stellar rest frame, for the 2019B (left) and 2020B (right) spectra of AU Mic. In both periodograms, the dashed horizontal line traces Prot and the orbital periods of the transiting and candidate planets, whereas the vertical dotted lines depict the velocities of the three components of the He I triplet. The color scale traces the logarithmic power in the periodograms. We caution that only the main peaks (colored yellow to red and extending over at least several velocity bins) are likely to be significant in these plots.

In the text