There is a newer version of the record available.

Published November 4, 2025 | Version v0

Foundation Models for Discovery and Exploration in Chemical Space

Description

Source Code and Data Drop for Foundation Models for Discovery and Exploration in Chemical Space.

File Contents
electrolyte-fm-main.zip Source code for MIST.
MODELS.md Specification for all model checkpoints.
mist-pretrained.tar.gz Checkpoints (in Safetensors format) for pretrained variants of the MIST-28M and MIST-1.8B encoder.  Denoted by mist-{size}-{run_id} in MODELS.md. README in each models directory contains further usage instructions and examples.
mist-28M-finetuned.tar.gz Checkpoints (in Safetensors format) for finetuned variants of the MIST-28M encoder used to evaluate benchmark performance and demonstrate downstream applications. Denoted by mist-28M-{run_id}-{dataset} in MODELS.md. README in each model directory contains further usage instructions and examples.
mist-1.8B-molnet-xx.tar.gz

Checkpoints (in Safetensors format) for finetuned variants of the MIST-1.8B encoder used to evaluate MoleculeNet benchmark performance (except QM9). Models are denoted by mist-1.8B-{run_id}-{dataset} in MODELS.md. README in each model directory contains further usage instructions and examples.

  • mist-1.8B-molnet-0.tar.gz : contains models for datasets HIV, BBBP and ESol.
  • mist-1.8B-molnet-1.tar.gz : contains models for datasets FreeSolv, Lipophilicity and SIDER.
  • mist-1.8B-molnet-2.tar.gz : contains models for datasets BACE, ClinTox and Tox21.
mist-1.8B-qm9-xx.tar.gz

Checkpoints (in Safetensors format) for finetuned variants of the MIST-1.8B encoder used to evaluate QM9 benchmark performance. Models are denoted by mist-1.8B-{run_id}-{target} in MODELS.md. README in each model directory contains further usage instructions and examples.

  • mist-1.8B-qm9-0.tar.gz : contains models for target properties G298, H298 and ZPVE.
  • mist-1.8B-qm9-1.tar.gz : contains models for target properties U298, U0, mu.
  • mist-1.8B-qm9-2.tar.gz : contains models for target properties Cv, HOMO and Gap.
  • mist-1.8B-qm9-3.tar.gz : contains models for target properties alpha, r2, and LUMO.
mist-mixtures.tar.gz Checkpoints (in Safetensors format) for finetuned mixture property prediction models. Denoted by mist-conductivity-27.0M-2mpg8dcd and mist-mixtures-zffffbex in MODELS.md. README in each model directory contains further usage instructions and examples.
mixtures.tar.xz   Training and validation datasets (including experimental validation), as well as calculated SOAP fingerprint similarity values for excess property models. 
screening.tar.xz Logs and databases with generated molecules for all screening runs are in the runs directory. SMILES encodings and predicted properties for the reference set and Pareto front candidates are in electrolytes_predictions.csv and generate_pareto_front.csv respectively.
linear_probes.tar.xz   Dataset as well as weights, statistics and metadata for linear probes used to identify Lipinski Rule of Five features (serialized using JLD2).
scaling.tar.xz MCMC chains, including posterior estimates for all
model parameters, and convergence metrics for all scaling models tested (serialized using JLD2).
run_logs.tar.xz

All hyperparameter settings, loss curves, system utilization and throughput measurements for all training runs, including model development and production runs.

 

Files

electrolyte-fm-main.zip

Files (84.9 GB)

Name Size
md5:ec2871b2f78986ac3df6c85e38a3acec
685.9 kB Download
md5:79690645076cd3c7caf836ec70e856f6
1.6 MB Preview Download
md5:6974bde0693b4f013070512e7a1c1049
35.3 MB Download
md5:f48cb7a9557b84990ce99e28c091f425
10.2 GB Download
md5:ed228764e91a41e0107c7bf69e1a93eb
16.8 GB Download
md5:707aaeeb7b5dbbc4aab0b516c2a0ea23
16.8 GB Download
md5:d712172c9e4289eb30e0bfe10cd8fc18
7.8 GB Download
md5:256d074f4ecacd21d15f739275c7ccf9
7.8 GB Download
md5:bc619ace4e4d141a9ccb10fd574be8d5
7.8 GB Download
md5:f3c8b00b854753c643de37756b2b86a0
10.1 GB Download
md5:b3ae99d2b2b4c88a62d58a712ff95682
1.9 GB Download
md5:3573069abe432692b302eb4bdaa9cbb3
201.5 MB Download
md5:9bebc3aedd92880264b1aecb3842cb12
3.5 GB Download
md5:ab43d843a7cfc163c8fd741a32eb4241
21.5 MB Download
md5:e3cf813723870a760429d578a497adb0
7.3 kB Preview Download
md5:11f7522707c417063d6a3436b87dd3b3
56.0 MB Download
md5:62ce2a59a5652a6d2bde716d17bb355a
19.4 MB Download
md5:34553528cbdcb6836cf65d15784dca20
1.9 GB Download

Additional details

Related works

Is supplement to
Publication: arXiv:2510.18900 (arXiv)

Software

Repository URL
https://github.com/BattModels/mist
Programming language
Python , Julia
Development Status
Wip

References

  • Wadell, Alexius and Bhutani, Anoushka et al. "Foundation Models for Discovery and Exploration in Chemical Space." arXiv:2510.18900. Preprint, arXiv, October 20, 2025. https://doi.org/10.48550/arXiv.2510.18900.