Chemistry-Informed Transfer and Interaction-Resolved Learning Towards Data-Efficient Materials Exploration

Optimizing synthesis parameters for new thin-film materials is slow and often relies on intuition and trial and error. Gaussian process (GP) models suit low-data settings, but treat each material separately or need engineered descriptors to learn across materials. Multi-task GPs share information across materials but infer these relationships solely from sparse data, making them unreliable early on. We introduce a chemistry-informed multi-task GP that derives these relationships from a Tanimoto kernel on molecular fingerprints, enabling knowledge transfer without descriptor engineering. In closed-loop, uncertaintydriven exploration of a seven-dimensional synthesis space, comprising one categorical variable (three chiral 2D hybrid organic-inorganic perovskite compositions) and six continuous synthesis parameters, we tested 287 conditions (<0.1% of the discretized synthesis space). Knowledge transfer matched the accuracy of single-task GPs trained on the full target dataset using between 35-60% of the target data. Interaction analysis revealed a three-way coupling among annealing temperature, precursor concentration, and composition. This coupling makes a one-variable-at-a-time optimization challenging. Our framework offers a simple strategy for transferring and interpreting knowledge across related materials.