This repository contains the core generative modeling pipeline for property-conditioned polymer discovery, leveraging a Conditional Discrete Diffusion Language Model (CDDLM) built on a modern transformer encoder backbone.
Unlike traditional autoregressive sequence models, this implementation frames molecular generation as a non-autoregressive, iterative denoising process. The model learns to recover clean polymer SMILES structures from completely corrupted (masked) sequences, explicitly guided by continuous physical property constraints.
- Generative Backbone: Utilizes answerdotai/ModernBERT-base as a bidirectional transformer encoder to capture dense, long-range contextual sequence representations across complex macromolecular architectures.
- Property Conditioning: Implements a continuous GaussianFourierProjection embedding module that maps scalar property constraints—such as target electronic band gap (E_g)—into a high-dimensional frequency space to drive the reverse diffusion trajectory.
- Sampling Strategy: Incorporates Classifier-Free Guidance (CFG) during the reverse denoising process, enabling precise control over the trade-off between target property adherence and sequence diversity.
- Validation Pipeline: Features an integrated evaluation suite calculating standard structural metrics (Validity, Uniqueness, Novelty) alongside real-time quantum-chemical verification via a live GFN2-xTB calculator.
Evaluated on a pool of 10,000 raw generations, unconditioned by downstream MCTS search or filtering.
| Metric | Value |
|---|---|
| Target property (E_gc) | 4.50 eV |
| Generative pool size | 10,000 |
| Chemical validity rate | 3.94% |
| Uniqueness rate | 64.72% |
| Novelty rate | 100.00% |
| Internal diversity (IntDiv) | 0.7236 |
| Target property adherence | 0.4105 eV (MAE) |
Note on validity: the 3.94% raw validity rate reflects unfiltered, unguided diffusion
output — this repo covers Phase 1 of the pipeline only. This low raw rate is the direct
motivation for the MCTS-guided search layer (polymer_mcts_diffusion), which reranks and
filters generations toward chemical validity and target-property adherence rather than
relying on the base model alone. The high novelty (100%) and moderate-to-high uniqueness
(64.72%) indicate the model is exploring genuinely new chemical space rather than
memorizing or collapsing to a narrow mode — the problem the search layer is built to
capitalize on.
Status: Phase 2 (offline critic training) and Phase 3 (online MCTS-guided RL)
evaluation in progress — see polymer_mcts_diffusion and polymer_mcts_transformer for
architecture; quantitative results forthcoming there.
- model.py — Core CDDLM architecture and GaussianFourierProjection embedding layers.
- tokenizer.py — Vocabulary mappings and regex-based tokenization optimized for handling complex polymer branching and wildcards (* / [*]).
- training.py — Main unconditioned/conditioned pre-training script utilizing the PI1M_v2.csv dataset.
- finetune_training.py — Fine-tuning pipeline built for strict conditioning on explicit electronic properties via Egc.csv.
- evaluate_metrics.py — Standardized validation matrix computing internal metrics alongside external quantum-chemical property adherence via the live xTB calculator.
- finetune_inference.ipynb — Interactive workspace for checkpoint evaluation, diverse sample stream generation, and property verification.
- train.sh — Pre-training execution script.
- finetune.sh — Fine-tuning execution script.
The pipeline expects paths to two primary data tracking files:
PI1M_v2.csv: Large-scale polymer database used for capturing structural syntax and baseline synthetic accessibility profiles.Egc.csv: Target dataset containing explicit SMILES mappings to calculated electronic band gap (E_g) values.
To run baseline pre-training: python training.py (or run: bash train.sh)
To execute property-conditioned fine-tuning for targeted electronic profiles: python finetune_training.py (or run: bash finetune.sh)
This codebase is a component of ongoing graduate research at the Georgia Institute of Technology (School of Materials Science & Engineering).
This project is licensed under the MIT License - see the LICENSE file for details. © 2026 Vansh Suresh Yadav. All rights reserved. This code is intended exclusively for private research evaluation. Copying, distributing, or modifying these files without explicit authorization is strictly prohibited.