More Data or Better Data? Conditioned Cotraining of Spin-Aware Foundation Models for Heterogeneous Catalysis

The rapid evolution of machine learning interatomic potentials (MLIPs) has promised a new era of high-throughput catalyst discovery. However, universal foundation models often suffer from critical "blind spots" that hinder their utility in industrial heterogeneous catalysis. These models typically rely on datasets consisting of pristine crystal facets at 0 K, failing to account for the electronic complexity, spin polarization, and chemical diversity inherent in real-world chemical reactors. In this talk, we present a systematic approach to fixing these blind spots using targeted data and multi-fidelity fine-tuning strategies.We will discuss three specialized datasets designed to bridge these gaps: AQCat25, AQVolt26, and AQCat26. AQCat25 provides spin-aware density functional theory (DFT) calculations to address magnetic transition metal systems where electronic effects govern activity. AQVolt26 captures the anharmonic lattice dynamics and soft lattice fluctuations often ignored by standard equilibrium-based models but vital for high-temperature stability. AQCat26 expands the accessible reaction space to encompass unique adsorbates with chain lengths up to C8, providing the chemical diversity necessary for industrial modeling. Together, these datasets enable the investigation of complex reaction networks and multi-carbon product synthesis that remain beyond the reach of existing spin-agnostic foundation models.We evaluate integration strategies for these data streams, including multi-fidelity co-training with feature-wise conditioning and direct fine-tuning. Our results demonstrate that co-training on a mixture of universal and targeted data allows models to map extreme physical regimes—including defects and vacancies—without inducing catastrophic forgetting of the general chemical space. This methodology yields significant improvements in force prediction accuracy and captures essential scaling relations for thermochemical reaction networks. This work shows that achieving industrial relevance requires the strategic injection of targeted physics to move beyond the limitations of model scale alone.