Disclaimer: This is a snapshot of an actively developed project, not the latest version. Ongoing work continues in a private repository and has not been synced here. Contact Dr. Mohsen Rezayat for more details: mohsen.rezayat@omid-usa.org
A hardened, DFT-validated extension of CycleChemist, a dual-pronged machine learning framework for organic photovoltaic (OPV) discovery. The original framework paired a generative molecular transformer (MatGPT) with GNN property predictors (MOE2/P3, OPVC) and a reinforcement learning loop to design novel donor:acceptor (D:A) molecular pairs. This version rebuilds that RL loop around real quantum-chemistry validation and reward-hacking-resistant gating, closing the gap between what the reward model scores highly and what is actually a synthesizable, spectrally plausible D:A candidate.
The base CycleChemist reward signal (PCE and transparency scores from a learned surrogate) is fast but exploitable: a policy can learn to maximize the surrogate's score without producing chemistry that a domain expert, or a real quantum-chemical calculation, would accept. This version closes that gap with a three-phase gating system layered directly into the RL fine-tuning loop (matgpt/code/finetune.py):
-
Phase 1: Structural plausibility gate (
matgpt/code/structural_gates.py) — scores every RL-generated candidate against the literature-typical D-A conjugated-polymer motif (extended conjugation, a solubilizing alkyl chain, donor/acceptor SMARTS motifs) using soft, multiplicatively-combined checks, so a candidate cannot earn full credit by satisfying only one axis of "looks like a real OPV polymer." -
Phase 2: Curated D-A fragment library and synthetic corpus builder (
matgpt/code/fragment_library.py,matgpt/code/build_fragment_corpus.py) — assembles donor and acceptor building blocks, each tagged with RDKit attachment points, into synthetically-plausible D-A oligomers viamolzip, with randomized solubilizing-chain lengths per sample. This corpus is used to fine-tune MatGPT toward the real D-A polymer chemical space before RL ever begins. -
Phase 3: Oracle-driven validation gate (
property_predictors/moe2_p3/dftb_oracle.py,phase3_validate.py,validate_top_candidates.py) — a DFTB+/xTB pipeline computes actual TD-DFTB excitation spectra for candidate molecules and derives Phi_vis / AVT from them, giving the RL loop a ground-truth-adjacent signal instead of relying solely on a learned PCE surrogate. The gate is triggered selectively during RL fine-tuning, renormalizes and calibrates the spectral surrogate's output against oracle spectra, and gates the spectral reward path by OPV-material probability so a high transparency score can no longer be earned by drifting outside the OPV-relevant chemical space. Known TD-DFTB limitations (systematic underestimation of charge-transfer excitation energies in donor-acceptor push-pull systems) are documented in-code rather than silently absorbed into the reward. -
RL training stability fixes: gradient clipping and a floored reward-anneal term in the fine-tuning loss, plus a pass of CLI argument fixes (
store_truefootguns) that had been silently producing wrong runs. -
Test coverage:
structural_gates.pyandfragment_library.pyship with pytest suites (matgpt/code/tests/) — the RL reward's gating logic has regression coverage where the original codebase had none. -
Repo hygiene: checkpoint directories are untracked and gitignored rather than committed, keeping the repo lightweight and reproducible from config instead of from stored weights.
Net effect: the generation-validation loop no longer trusts a single learned scalar. A candidate has to pass a structural plausibility check, come from (or resemble) a corpus of synthetically real D-A chemistry, and hold up under an actual excited-state quantum-chemistry calculation before it is rewarded as a good OPV candidate.
The underlying discovery framework combines material generation, property prediction, and reinforcement learning for iterative improvement:
This work builds directly on CycleChemist, originally developed by Hou Hei Lam, Jiangjie Qiu, Xiuyuan Hu, Wentao Li, Fankun Zeng, Siwei Fu, Hao Zhang, and Xiaonan Wang — see CycleChemist: A Dual-Pronged Machine Learning Framework for Organic Photovoltaic Discovery and the earlier Accelerating High-Efficiency Organic Photovoltaic Discovery via Pretrained Graph Neural Networks and Generative Reinforcement Learning. The MatGPT generative model, the MOE2/P3 and OPVC property predictors, and the original RL fine-tuning scaffold are their work. The DFTB+/xTB oracle pipeline, spectral calibration, multi-phase reward gating, fragment-based corpus construction, and associated test coverage described above are additions in this version, maintained by Om (@OJ102).
- Experimental dataset (used in experiments):
data/exp_dataset.csv - Active / continuously updated dataset: OPV2D
- D-A fragment library (this version):
data/fragments/da_fragments.csv
This repo was developed with Python 3.9. A minimal setup is:
conda create -n CycleChemist python=3.9 -y
conda activate CycleChemist
# PyTorch (example: CUDA 11.8; adjust for your machine)
pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu118
# Core scientific stack
conda install -c conda-forge numpy pandas tqdm psutil scikit-learn rdkit torch-geometric tensorboard matplotlib seaborn -yTo use the DFTB+/xTB oracle pipeline (Phase 3), also install DFTB+ with the 3ob-3-1 Slater-Koster parameter set (bundled under DFTB/) and set DFTB_PARAM_DIR to its location.
cd data/
wget ftp://ftp.ncbi.nlm.nih.gov/pubchem/Compound/Extras/CID-SMILES.gz
gunzip CID-SMILES.gz
tr '\t' ',' < CID-SMILES > CID-SMILES.csv
sed -i '1i CID,SMILES' CID-SMILES.csv
python down_sample.py
cd ..Source: "Design Principles and Top Non-Fullerene Acceptor Candidates for Organic Photovoltaics" (https://www.sciencedirect.com/science/article/pii/S2542435117301307)
cd data/
wget -O mmc2.csv "https://ars.els-cdn.com/content/image/1-s2.0-S2542435117301307-mmc2.csv"
cd ..python matgpt/code/pretrain.py --config matgpt/pretrain_config/pubchem.json
tensorboard --logdir matgpt/runs/matgpt_pretrain_pubchem/python property_predictors/opvc/build_opv_dataset.py
python property_predictors/opvc/train_rf_classifier.pypython property_predictors/moe2_p3/pretrain.py
python property_predictors/moe2_p3/train_pce.py
tensorboard --logdir property_predictors/moe2_p3/runs/MLM-Pretrain
tensorboard --logdir property_predictors/moe2_p3/runs/HOMO-Finetune
tensorboard --logdir property_predictors/moe2_p3/runs/HOMO-EXP-Finetune
tensorboard --logdir property_predictors/moe2_p3/runs/PCE-Predict# Fit to data/exp_dataset.csv
python matgpt/code/pretrain.py --config matgpt/pretrain_config/exp_dataset.json
tensorboard --logdir matgpt/runs/matgpt_pretrain_exp_dataset/Build a synthetic D-A oligomer corpus from the curated fragment library and fine-tune MatGPT toward it before RL:
python matgpt/code/build_fragment_corpus.py
python matgpt/code/pretrain.py --config matgpt/pretrain_config/hendsbee_da_corpus.jsonCase study: fix acceptor (Y6) and search for donors
python matgpt/code/finetune.py \
--run_name donor_Y6 \
--prior_path matgpt/ckpt/matgpt_pretrain_exp_dataset/final_model.pt \
--target_type "donor" \
--fixed_smiles "CCCCCCCCCCCc1c(/C=C2\C(=O)c3cc(F)c(F)cc3C2=C(C#N)C#N)sc2c1sc1c3c4nsnc4c4c5sc6c(CCCCCCCCCCC)c(/C=C7\C(=O)c8cc(F)c(F)cc8C7=C(C#N)C#N)sc6c5n(CC(CC)CCCC)c4c3n(CC(CC)CCCC)c21" \
--use_homo_lumo_loss
tensorboard --logdir matgpt/runs/donor_Y6Case study: fix donor (PTB7-Th) and search for acceptors
python matgpt/code/finetune.py \
--run_name acceptor_PTB7-Th \
--prior_path matgpt/ckpt/matgpt_pretrain_exp_dataset/final_model.pt \
--target_type "acceptor" \
--fixed_smiles "CCCCC(CC)COC(=O)c1sc2c(-c3cc4c(-c5ccc(CC(CC)CCCC)s5)c5sc(C)cc5c(-c5ccc(CC(CC)CCCC)s5)c4s3)sc(C)c2c1F" \
--use_homo_lumo_loss
tensorboard --logdir matgpt/runs/acceptor_PTB7-ThBoth case studies now run against the Phase 1/2/3 gates by default when the corresponding config flags are set — see matgpt/code/finetune.py for the structural-gate, oracle-trigger, and reward-anneal flags.
Validate a trained spectral surrogate, or a batch of top RL candidates, against real DFTB+/TD-DFTB spectra:
python property_predictors/moe2_p3/phase3_validate.py --n-workers 4
python property_predictors/moe2_p3/validate_top_candidates.pypython matgpt/code/sample.py \
--model_path matgpt/ckpt/rl_fine_tune/donor_Y6/step500.pt \
--vocab_path data/vocab_pubchem_len20-290_no_ions_no_multi_random5M_123.txt \
--output_path results/out.csvThere is yet to be a published version of this repository. If you use this work in your research, please cite the repository and the original CycleChemist papers:
@misc{qiu2025,
title={Accelerating High-Efficiency Organic Photovoltaic Discovery via Pretrained Graph Neural Networks and Generative Reinforcement Learning},
author={Jiangjie Qiu and Hou Hei Lam and Xiuyuan Hu and Wentao Li and Siwei Fu and Fankun Zeng and Hao Zhang and Xiaonan Wang},
year={2025},
eprint={2503.23766},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2503.23766},
}
@misc{lam2025cyclechemistdualprongedmachinelearning,
title={CycleChemist: A Dual-Pronged Machine Learning Framework for Organic Photovoltaic Discovery},
author={Hou Hei Lam and Jiangjie Qiu and Xiuyuan Hu and Wentao Li and Fankun Zeng and Siwei Fu and Hao Zhang and Xiaonan Wang},
year={2025},
eprint={2511.19500},
archivePrefix={arXiv},
primaryClass={cond-mat.mtrl-sci},
url={https://arxiv.org/abs/2511.19500},
}And cite the original data sources:
@article{Min2020,
author = {Wu, Yao and Guo, Jie and Sun, Rui and Min, Jie},
title = {Machine learning for accelerating the discovery of high-performance donor/acceptor pairs in non-fullerene organic solar cells},
journal = {npj Computational Materials},
volume = {6},
number = {1},
pages = {120},
year = {2020}
}
@article{Saeki2021,
author = {Miyake, Yuta and Saeki, Akinori},
title = {Machine Learning-Assisted Development of Organic Solar Cell Materials: Issues, Analyses, and Outlooks},
journal = {The Journal of Physical Chemistry Letters},
volume = {12},
number = {51},
pages = {12391-12401},
year = {2021}
}
@article{lopez2017design,
title = {Design principles and top non-fullerene acceptor candidates for organic photovoltaics},
author = {Lopez, Steven A and Sanchez-Lengeling, Benjamin and de Goes Soares, Julio and Aspuru-Guzik, Al{\'a}n},
journal = {Joule},
volume = {1},
number = {4},
pages = {857--870},
year = {2017},
publisher = {Elsevier}
}