Towards a Spin-Aware Foundation Potential for Heterogeneous Catalysis

Machine learning force fields (MLFFs) promise substantial acceleration of computational catalyst discovery compared to traditional Density Functional Theory (DFT) calculations. Despite this advantage, fundamental limitations persist in capturing the electronic structure complexity necessary for accurate transition metal catalysis predictions. Most existing MLFFs optimize for coverage across vast chemical spaces rather than preserving the magnetic and electronic properties that govern catalytic activity. This focus on breadth over electronic fidelity creates systematic errors in predicting spin-dependent processes, which are ubiquitous in transition metal systems where magnetic moments and spin states directly influence reaction pathways and energetics. This work addresses the critical gap through development of spin-aware MLFFs that incorporate magnetic effects essential for transition metal catalysis. We validate the MLFFs through testing on their capability for accurately capturing key scaling relations of representative thermochemical reaction networks, such as syngas to multi-carbon products and ammonia synthesis, as well their generalizability to more out-of-domain tasks such as the evaluation of spin-mediated promotion mechanisms. Finally, we test the MLFFs foundational propensity, by assessing its capability to reduce the number of spin-polarized DFT calculations necessary (often 5 to 20 times more costly than spin-unpolarized calculation) when expanding to compositionally distinct adsorbates than those in the model’s training set. The incorporation of magnetic effects represents a critical advancement toward physics-grounded AI-accelerated modeling tools capable of predicting fundamental mechanisms beyond the scope of existing spin-agnostic models.