1 of 32

CANSTAR:

Mitigating stellar radial velocity jitter

using orthogonal activity indices

and a time-aware neural network

​

Jordi Blanco-Pozo (blanco@ice.csic.es, jordiblp.github.io),

Manuel Perger, Guillem Anglada-Escudé, Ignasi Ribas,

David Baroch, Marina Lafarga, Juan Carlos Morales,

Òscar Porqueras-León, Sophie Stucki, and David Vallmanya Poch

Funded by Grant PRE2022-101942

Blanco-Pozo et al. 2026, A&A, 710, A194

2 of 32

The CCF is sensitive to Doppler shifts

But also to line shifts and line shapes induced by stellar activity

e.g. spots and faculae

3 of 32

time-series

spectroscopic observable

spectroscopic observable

spectroscopic observable

Convolutional layers

an coeff time-series

spectroscopic time-series

an coeff time-series

an coeff time-series

gi time-series

Nf filters

xNc convolutional layers

Multi-head attention

Flatten

Linear layer

RV time-series

Layer norm

Feed-forward layer

Layer norm

xNe encoder layers

Transformer encoder

4 of 32

time-series

spectroscopic observable

spectroscopic observable

spectroscopic observable

Convolutional layers

an coeff time-series

spectroscopic time-series

an coeff time-series

an coeff time-series

gi time-series

Nf filters

xNc convolutional layers

Multi-head attention

Flatten

Linear layer

RV time-series

Layer norm

Feed-forward layer

Layer norm

xNe encoder layers

Transformer encoder

Separating line shifts

from line shape distortions

with an orthonormal basis function

5 of 32

time-series

spectroscopic observable

spectroscopic observable

spectroscopic observable

Convolutional layers

an coeff time-series

spectroscopic time-series

an coeff time-series

an coeff time-series

gi time-series

Nf filters

xNc convolutional layers

Multi-head attention

Flatten

Linear layer

RV time-series

Layer norm

Feed-forward layer

Layer norm

xNe encoder layers

Transformer encoder

Correct for the Stellar Activity in RV data

with a Convolutional Attention Network

(agnostic to the RV data)

6 of 32

Separating line shifts from line shape distortions

with an orthonormal basis function

line shifts

line shape distortions

7 of 32

Separating line shifts from line shape distortions

with an orthonormal basis function

line shifts

line shape distortions

8 of 32

Separating line shifts from line shape distortions

with an orthonormal basis function

line shifts

line shape distortions

9 of 32

Separating line shifts from line shape distortions

with an orthonormal basis function

line shifts

line shape distortions

10 of 32

CANSTAR:

Convolutional AttentioN for Stellar Activity Removal

time-series

spectroscopic observable

spectroscopic observable

spectroscopic observable

Convolutional layers

an coeff time-series

spectroscopic time-series

an coeff time-series

an coeff time-series

gi time-series

Nf filters

xNc convolutional layers

Multi-head attention

Flatten

Linear layer

RV time-series

Layer norm

Feed-forward layer

Layer norm

xNe encoder layers

Transformer encoder

Linear layers

spectroscopic observable

spectroscopic observable

spectroscopic observable

Scaled-dot

product attention

xNh heads

Linear layers

Linear layers

Linear layer

Flatten

Multi-head attention

xNh heads

11 of 32

conv attention network

time-series

spectroscopic observable

spectroscopic observable

spectroscopic observable

Convolutional layers

an coeff time-series

spectroscopic time-series

an coeff time-series

an coeff time-series

gi time-series

xNf filters

xNc convolutional layers

Multi-head attention

Flatten

Linear layer

RV time-series

Layer norm

Feed-forward layer

Layer norm

xNe encoder layers

Transformer encoder

CANSTAR: Convolutional AttentioN for Stellar Activity Removal

Convolutional Attention Network models the line-shape distortions, without ever seeing the line-shift (the RV!)

12 of 32

StarSim 3

Flux, suppression of convection, limb darkening

​

  • Photometry

​

  • Spectroscopy
    • Transmission spectroscopy
    • CCF

​

Stellar models: PHOENIX

​

Herrero et al. (2016), Rosich et al. (2020),

Stucki et al. (in prep.), Gomes et al. (in prep.)

Gomes, Fernandez-Puig et al. (in prep.)

https://github.com/dbarochlopez/starsim

13 of 32

Eps Eri

Young (400-800 Myr) nearby K2 dwarf with a rotation period of 11.2 days.

​

Eps Eri b, Jupiter-mass exoplanet on a 7.3 years.

​

HARPS 2019 observing run to study stellar activity and set constraints on any shorter period exoplanet.

​

​

14 of 32

TZ Ari

Nearby M5 dwarf with a rotation period of 1.96 days.

​

TZ Ari b, Saturn-mass exoplanet on a 2.12 year orbit.

​

CARMENES 2016-2019 observing run which samples both the long period planet and the short period stellar activity signal.

​

​

15 of 32

ε Eridani

TZ Arietis

16 of 32

Training on simulated StarSim data

ε Eridani

TZ Ari

17 of 32

Evaluation on ε Eridani HARPS data

CANSTAR

observations

ε Eri b

RMS residuals 2.46 m/s, 52.5%

18 of 32

Evaluation on TZ Ari CARMENES data

1.96 d (Prot)

772 d (planet)

keplerian

CANSTAR

observations

RMS residuals 6.71 m/s, 62.4%

19 of 32

Evaluation on TZ Ari CARMENES data

1.96 d (Prot)

keplerian

CANSTAR

observations

RMS residuals 3.89 m/s, 33.4%

20 of 32

CANSTAR: Performance

ε Eridani: performance limited by the synthetic gap

​

TZ Ari: performance limited by photon noise (depending on the observation strategy)

​

​

Synthetic data with noise

Real data

ε Eridani

12.9%

51.1%

TZ Ari

35.7%

62%, 33%

21 of 32

SunSim + CANSTAR

Mitigating solar radial velocity jitter

using the PLATO RV Data Challenge dataset as benchmark

​

Jordi Blanco-Pozo, Sophie Stucki, Ignasi Ribas, Nina-Elizabeth Nèmec, Manuel Perger, Sowmya Krishnamurthy, Alexander Shapiro and collaborators

22 of 32

SunSim

StarSim spinoff code, benchmarked against Solar data (Stucki et al. in prep.).

​

3D MHD stellar models from MURaM MPS-ATLAS (Vögler et al. 2005, Kostogryz et al. 2024, Witzke et al. 2024).

​

Outstanding results reproducing the variability of 10 years HARPS-N Solar data close to instrumental precision and TSI photometric variability

​

More details in Sophie Stucki’s talk on Thursday!

23 of 32

SunSim: forward modelling SDO images

24 of 32

SunSim+CANSTAR:

Sun-as-an-exoplanet host star

25 of 32

SunSim

CANSTAR

26 of 32

PLATO RV Data Challenge

Cretignier et al. (2026)

27 of 32

Very Preliminary Results

CANSTAR

observations

28 of 32

Very Preliminary Results

29 of 32

This is only the beginning…

With a larger training set, with noise injected into the simulations and hyperparameter optimization of the network, we can expect to get to even better stellar activity corrections.

​

But we are already at the sub-m/s level of stellar activity correction, minimizing the synthetic gap.

​

Simulations are for a G2 star: we can extend the correction to other Sun-like stars.

​

​

30 of 32

Conclusions

We have presented a novel method which allows:

  • To separate line shifts from line shapes distortions, allowing to recover the classical activity indicators but also getting to higher order line shape changes.
  • Use this information as input to a Convolutional Attention network, agnostic to the RV time-series:
    • Model stellar activity on time domain
    • Importance of simulating stellar activity but also the photon noise.

​

Next step: SunSim + CANSTAR project.

  • Preliminary results: sub-m/s level of correction on PLATO RV Data Challenge dataset.
  • CCFs with other masks in order to maximize stellar activity signals.
  • Extend framework to other Sun-like stars.

​

Blanco-Pozo et al. 2026, A&A, 710, A194

31 of 32

Backup slides

Funded by Grant PRE2022-101942

32 of 32

StarSim data generation

Stellar Parameters

Spot and facula map