ReactFlow: Pretrained Checkpoints for Transition State and Reaction Pathway Generation

Abstract

ABOUT ReactFlow is a flow-matching model that generates the transition state (TS) and the surrounding reaction pathway from reactant and product structures. It uses an SO(3)-equivariant EquiformerV2 backbone (35.3M parameters) and is trained in two stages: flow-matching pretraining, followed by GRPO reinforcement-learning post-training against an energy reward. Code: https://github.com/Yifang-Qin/ReactFlow FILES reactflow_t1x_pretrain.tar.gz Transition1x, flow-matching pretraining reactflow_t1x_grpo.tar.gz Transition1x, GRPO post-training reactflow_halo8_pretrain.tar.gz Halo8, flow-matching pretraining reactflow_halo8_grpo.tar.gz Halo8, GRPO post-training SHA256SUMS SHA-256 checksums of the four archives Each archive unpacks to a single directory: {name}/hparams.yaml full training configuration {name}/checkpoints/last.ckpt model weights (PyTorch state_dict) The pretraining checkpoints include both the raw and the EMA weights. The GRPO checkpoints are initialized from the pretraining checkpoint of the same dataset. The files contain model weights only (no optimizer state), so they are intended for sampling, evaluation, and as initialization for post-training, not for resuming an interrupted run. TRAINING DATA Transition1x: reference level of theory wB97X/6-31G(d); GRPO energy reward from ANI-2x.Halo8: reference level of theory wB97X-3c; GRPO energy reward from AIMNet2. Both datasets are used in a compact 10-frame form (reactant, 8 intermediate NEB images, product) with an 80/10/10 train/validation/test split. The preprocessing scripts are provided in the code repository. USAGE git clone https://github.com/Yifang-Qin/ReactFlow && cd ReactFlow mkdir -p outputs tar -xzf reactflow_t1x_grpo.tar.gz -C outputs bash scripts/eval_t1x.sh See the repository README for installation, data preparation, and the evaluation of the other checkpoints.