Researchers from Microsoft Research AI for Science and Freie Universität Berlin developed Orbformer, a transferable wavefunction model designed to simulate chemical bond breaking, according to a study published in Nature Communications. The architecture applies Quantum Monte Carlo with deep neural networks to solve the multireference electronic structure challenges found in dissociating molecular species.
The authors pretrained Orbformer on 22,000 equilibrium and dissociating structures, enabling the model to be fine-tuned on previously unseen molecules. Across established benchmarks, bond dissociations, and Diels–Alder reactions, Orbformer was the only method tested that consistently converged to chemical accuracy of 1 kcal/mol. This design amortizes the computational expense of solving the Schrödinger equation over multiple systems rather than recalculating full electronic structures from scratch for every molecule.
Adam Foster, Zeno Schätzle, P. Bernát Szabó, and Lixue Cheng contributed equally to the paper as lead authors, collaborating with Jonas Köhler, Gino Cassella, Nicholas Gao, Jiawei Li, Frank Noé, and Jan Hermann. The authors acknowledged project funding from Microsoft Research, leadership from C. Bishop at AI for Science, quantum chemistry discussions with P. Gori-Giorgi, and GPU cluster management support from M. Riechert, T. Vogels, and H. Schulz.
Nature Communications accepted the peer-reviewed paper on July 31, 2026, following its initial submission on October 2, 2025, and published it on August 21, 2026. Author affiliations and present addresses listed in the publication include the Cambridge, UK laboratory of Microsoft Research, Freie Universität Berlin, Imperial College London, Tsinghua University, CuspAI, and The Hong Kong University of Science and Technology.
