Dynamic language model representations for multi-objective reaction optimisation

📄 Paper:

---

🔍 Overview

We present a general framework for multi-objective reaction optimisation that requires no descriptor computation or feature engineering. Textual descriptions of reaction conditions — ligands, additives, solvents, catalysts — are encoded by a fine-tuned language model and passed through objective-specific projection heads into independent surrogate models, jointly trained end-to-end. The representation adapts to the reaction system and objectives of the optimisation campaign at hand.

• Unifying chemically heterogeneous components, out of the box. A single representation can span mono- and bidentate phosphines, chiral Ir and Ru catalysts across different metals and coordination geometries, and additives from elemental zinc to organic and inorganic bases.
• Multiple objectives. Yield, selectivity, diastereomeric ratio, and enantiomeric excess can be optimised jointly.
• Low-data to high-throughput. Validated from sequential single-experiment campaigns to 24/48/96-well plate HTE regimes.
• Prospective validation. Two automated HTE campaigns reached process-relevant conditions in two rounds (192 reactions, <3% of each design space), translating directly to gram scale.

---

📦 Installation

Requires Python 3.10+.

git clone https://github.com/schwallergroup/alice.git
cd alice

Three dependency files are provided:

File	Use
environment.yml	Minimal conda environment — start here
requirements.txt	The same minimal set-up, for pip
environment.lock.yml	Transitive freeze of the environment used in this work (linux-64)

Quick start (conda):

conda env create -f environment.yml
conda activate alice
pip install -e .

Quick start (pip):

conda create -n alice python=3.10 -y
conda activate alice
pip install -r requirements.txt
pip install -e .

Exact environment used in this work (linux-64 only):

conda env create -f environment.lock.yml
conda activate alice
pip install -e .

All experiments in the manuscript were run on NVIDIA A100 GPUs with torch 2.8.0 (CUDA 12.8).

---

🧪 Running experiments

Retrospective benchmarks

The configs/ directory contains the configuration files used for the experiments in the manuscript, one per reaction dataset and representation:

configs/
├── ni_suzuki/             # Ni-catalysed Suzuki coupling
├── pd_suzuki/             # Pd-catalysed Suzuki coupling
└── pd_sulfonamide/        # Pd-catalysed sulfonamide C–N coupling

Each dataset provides three representations:

Config	Representation
<name>_desc.yaml	Numerical descriptors from descriptor libraries + OHE
<name>_ohe.yaml	One-hot encoding
<name>_llm.yaml	Fine-tuned LLM embeddings

Launch a single run with:

python train_master.py --config configs/<dataset>/<config>.yaml

W&B cloud tracking can be disabled by using (this still writes the local result CSV files and plots):

WANDB_MODE=disabled python train_master.py --config configs/<dataset>/<config>.yaml

The experiments in the manuscript were run as 20-seed W&B sweeps in configs/:

configs/
├── sweep_ni_suzuki.yaml
├── sweep_pd_sulfonamide.yaml
└── sweep_pd_suzuki.yaml

After logging in to W&B, initialize a sweep from the repository root using:

# to switch between featurisation methods, change parameters.config.value in the sweep file.
wandb sweep configs/<sweep_file>.yaml
# W&B will then print a sweep identifier of the form below which you can use to start the run:
wandb agent <entity>/<project>/<sweep-id>

---

🔬 Data

The data/ directory holds the three retrospective benchmark datasets used for evaluation in the manuscript, and the experimental data collected in the two prospective reaction optimisation campaigns.

data/
├── benchmark_datasets/
│   ├── ni_suzuki/             # Ni-catalysed Suzuki coupling
│   ├── pd_suzuki/             # Pd-catalysed Suzuki coupling
│   └── pd_sulfonamide/        # Pd-catalysed sulfonamide C–N coupling
└── prospective_campaigns/     # results of the two prospective optimisation campaigns

The two datasets ni_suzuki and pd_suzuki are measured HTE plates covering a designed subset of their parameter space; pd_sulfonamide is a virtual dataset enumerating the full grid.

Retrospective benchmarks

Each dataset ships the same reaction data in three representations, so that the featurizers can be compared on identical data:

File	Representation
<dataset>_desc.csv	Numerical descriptors from descriptor libraries + OHE
<dataset>_ohe.csv	One-hot encoding
<dataset>_prompt.csv	Natural-language reaction prompt, consumed by the LLM featurizers

Prospective campaigns

Results of the two prospective optimisation campaigns, in SURF format — one row per reaction, with CAS numbers and SMILES for every component:

File	Reaction	Objectives
asymmetric_hydrogenation_surf_reaction_data.csv	Asymmetric hydrogenation	conversion, de (syn), ee (syn)
pd_catalysed_cyanation_surf_reaction_data.csv	Pd-catalysed cyanation	conversion, selectivity

Each file contains 192 reactions — two 96-well HTE BO rounds, indexed by the BO round column.

---

Citation

@misc{sin2026dynamic,
  title         = {Dynamic language model representations for multi-objective reaction optimisation},
  author        = {Sin, Joshua W. and Segura, David Ming and Ranković, Bojana and Chau, Siu Lun and Lutz, Marius D. R. and Anelli, Andrea and Burwood, Ryan P. and Püntener, Kurt and Notheis, Maximilian J. and Bigler, Raphael and Schwaller, Philippe},
  year          = {2026},
  eprint        = {2609.11790},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  doi           = {10.48550/arXiv.2609.11790},
  url           = {https://arxiv.org/abs/2609.11790}
}

---

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.

---

Acknowledgements

This work was supported by NCCR Catalysis (grant no. 225147), a National Centre of Competence in Research funded by the Swiss National Science Foundation, and by the Swiss National Science Foundation (SNSF) [226509].