Action-Conditioned Sparse Mixture-of-Experts World Model for Predicting Microplate Observations in Robotic Chemistry
Automated chemistry requires both reaction-outcome prediction and monitoring of the physical execution of experiments. We study an action-conditioned model for predicting welllevel plate observations after robotic operations. The architecture combines address-masked action injection, sparse experts routed by a chemistry-oriented scene descriptor, and a gated pyramid decoder for local refinement. In the reported trajectory evaluation, the model reaches PSNR 29.6 and failure-detection F1 0.79 at 7.6 GFLOPs and 118 ms per plate. A separate evaluation on high-throughput reaction datasets reports yield 𝑅2 of 0.821. Because the available appearance and yield data come from different sources, these results assess complementary tasks rather than an end-to-end action–appearance–yield trajectory on the same plate. We also examine component ablations, cross-scenario behaviour and inference cost.
Authors
- Jiayi Ren
- Chenming Hao (ORCID: https://orcid.org/0009-0006-6249-6269)
- Zixuan Liao
- Yuxin Gu
- Tianhao Qiu
Publication Details
- Journal
- ChemRxiv
- Published
- 2026-11-16
- DOI
- https://doi.org/10.26434/chemrxiv.15012244/v1
- Primary Topic
- Machine Learning in Materials Science
- Type
- preprint