Inverse Model — Generative Design

Design
New Polymers

HAPPY + GPT steered via Reinforcement Learning to explore the Tg–Egb design space. Watch how generated structures evolve toward the target — then browse top candidates at any checkpoint.

Explore Trajectories →
Architecture

CI-LLM Inverse Model

GPT-based generative model trained on HAPPY sequences, steered via RL to target desired property ranges.

🎯

Target Properties

Specify desired Tg ≥ 600 K and Egb ≥ 4.5 eV simultaneously as RL reward targets.

🤖

GPT Generator

Autoregressive generation of HAPPY token sequences, constrained to valid polymer chemistries.

RL Fine-tuning

Reward: Tg, Egb, SAScore, Diversity (Tnms), Novelty — balancing property targeting and chemical diversity.

Pretrained GPT
Generate HAPPY
Forward Model (reward)
RL update
Better polymers
Interactive Explorer

Dual-Property Optimization

Simultaneously targeting Tg ≥ 600 K and Eg ≥ 4.5 eV via RL. Each point is the mean of 512 generated molecules at that step. Color gradient = training progress (light → dark). ★ = target.

Loading…
Top Molecules
ranked by RL reward · ±100 steps
Scaffold-Constrained RL

Imide-Targeted Generation

Chemistry-steered RL constrained to imide scaffold structures.

Loading trajectory…