SciMem: Scientific Reasoning with Structured Memory for Materials Design

Lee Hyun, Sohee Yoon, Jinwoo Park, Myeonghun Jeong

Track: Track 1: Original Research/Position/Education/Attention Track

Abstract: Large language models are increasingly used for scientific and engineering tasks, but domain adaptation often requires costly supervision or parameter updates. Memory-based methods offer a lightweight alternative by reusing past experience at inference time, which is especially attractive in materials science where experimental data are scarce and heterogeneous. Yet memory reuse in materials design is challenging because the effect of a condition depends on material class, target property, processing history, and measurement context, so naive accumulation can introduce irrelevant or conflicting experience. We propose SciMem, a structured memory framework for materials reasoning. SciMem represents experience as a Graph of Experience, whose nodes encode materials features and scientific effects and whose edges capture mechanistic dependencies. It further updates memory within disentangled scientific contexts such as mechanism family, material class, and target property. Experiments on property prediction and property-conditioned synthesis planning using Open Materials Guide show that SciMem improves reasoning performance and memory efficiency over prompting and unstructured-memory baselines.

Keywords: Reasoning, materials design, inverse design, LLM

Submission Number: 107