Abstract Electrochemical CO2 reduction provides a promising route to convert CO2 and renewable electricity into liquid fuels and chemicals, but selectively directing Cu-based catalysts toward specific C1 or C2 oxygenates remains challenging due to competing reaction pathways. A central challenge is to control not only the identity of Cu sites, but also their spatial connectivity and interfacial electronic environment that governs the fate of *CO-derived intermediates. In this work, we report that Cu loading on indium tin oxide (ITO) can serve as a systematic control parameter to tune Cu ensemble–oxide interface cooperativity in CO2 electroreduction and alter CO2RR pathways. By increasing Cu loading from 1 wt %, 5 wt % to 10 wt %, the dominant liquid product evolves from methanol (MeOH) to mixed methanol/ethanol (EtOH) and finally to EtOH. Importantly, ethylene (C2H4) was observed across the Cu/ITO series, while the EtOH/C2H4 ratio increases progressively with Cu loading, indicating that the loading-dependent change involves the branching of C2 intermediates after C–C bond formation as well. Structural and surface analyses reveal a concurrent evolution from highly dispersed Cu species toward larger Cu-rich domains, together with systematic changes in the Cu–ITO interfacial electronic environment. Together, these observations support an ensemble–interface cooperative picture in which increasing Cu loading promotes the development of Cu-rich ensembles while maintaining an electronically modified Cu–ITO interfacial environment, thereby shifting the liquid oxygenate pathway from MeOH toward EtOH. The concurrent increase in the EtOH/C2H4 ratio further suggests that these structural and interfacial changes also influence the fate of oxygen-containing C2 intermediates and the post C–C-coupling branching between C2H4 and EtOH. This work therefore provides system-specific insight into how Cu-domain structure and oxide-support interactions can redirect oxygenate pathways in Cu-catalyzed CO2 electroreduction.
Abstract Precision chemistry requires more than accurate prediction: it requires a traceable chain from measurement to representation, mechanism, design, and reproducible intervention. We argue that, within a declared operating envelope, a small set of physically meaningful coordinates and explicit relations can reduce multiscale prediction and design to a closed model that can be evaluated within specified error and computational budgets. We formulate five requirements for such coordinates─predictive sufficiency, dynamical closure, interventional stability, identifiability, and reachability─and organize them in a representation–closure–search–intervention (RCSI) cycle. Its central object is an input–output description mediated by physical variables whose relations can be examined and revised. Analytic, numerical, learned, and hybrid solvers are selected according to accuracy, cost, and latency. Uncertainty and provenance are retained at every stage, and models are evaluated by extrapolation, mechanism discrimination, and intervention success. Examples from catalytic interfaces, electronic and atomistic models, polymer field theory, microscopy, active matter, biomolecular design, and quantum-control planning illustrate both the opportunity and the limits. We conclude with prospective tests that can falsify each stage and a reporting standard for closed-loop scientific AI.
Abstract The direct partial oxidation of methane to methanol is an important catalytic challenge, and metal–organic frameworks (MOFs) with open transition-metal sites have emerged as promising candidates for this reaction. However, the complex electronic structure of their reactive intermediates often limits the reliability of standard density functional theory (DFT) predictions. In this work, we use fixed-node diffusion Monte Carlo (FNDMC) to establish accurate reaction energetics for the complete methane-to-methanol conversion pathway using two representative cluster models, Fe2–BTC(Mg) and Mn2–HTPA(Mg). These models represent an Fe-based MOF constructed from 1,3,5-benzenetricarboxylate (BTC) linkers and an Mn-based MOF constructed from tris((4-carboxyl)phenylduryl)amine (HTPA) linkers, respectively. FNDMC predicts mildly endothermic O atom adsorption on both systems, with adsorption energies of 0.66(5) and 0.71(4) eV for Fe2–BTC(Mg) and Mn2–HTPA(Mg), respectively. For Fe2–BTC(Mg), FNDMC identifies a septet ground state for the oxygen-bound intermediate, consistent with Fe(III)–oxyl character, whereas all tested DFT functionals favor a quintet state. FNDMC predicts that the C–H activation step is more exothermic for Fe2–BTC(Mg), with a reaction energy of −0.51(6) eV, than for Mn2–HTPA(Mg), for which it is only slightly exothermic at −0.16(6) eV. Methanol desorption is 0.97(7) eV more endothermic on Mn2–HTPA(Mg) than on Fe2–BTC(Mg), indicating stronger product retention despite their similar O atom adsorption energies. These results show that FNDMC captures differences in reaction energetics and electronic structure that standard DFT may miss, providing a robust reference for reactive MOF intermediates.
Abstract Recent AI systems can autonomously execute scientific workflows and empirical software tasks. We investigate whether an autonomous agent can use evaluation feedback to iteratively develop a scientific model. To address this question, we developed AutoResearch-MLIP, a research protocol in which the agent follows predefined instructions to propose, implement, and evaluate MLIP designs, with outcomes carried forward to guide subsequent iterations. Starting from a simple pair-distance MLIP, each generation selects a parent model by balancing exploration and exploitation, while structured design records guide subsequent proposals. Across 40 generations, the agent explores MLIP designs spanning local message passing, radial representations, scalar-vector states, body-order terms, residual readouts, long-range-inspired terms, and periodic image-offset graph construction. The final G40 model is competitive with established MLIP baselines in post hoc benchmarks. These results show that, within predefined rules, an autonomous agent can use evaluation feedback to iteratively improve MLIP implementations.
Abstract The recently proposed partial-active-space (PAS) multistate second-order perturbation theory (PASPT2) [Precision Chemistry 4, 997 (2026)] features connected amplitudes and a connected and closed intermediate Hamiltonian. Despite these hallmarks, PASPT2 (denoted as PASPT2H from now on) is not strictly size-extensive, as originally thought (and numerically confirmed), albeit strictly size-consistent. Nevertheless, PASPT2 is near-extensive for the states with major projections on the chosen PASM0. This becomes more transparent upon introducing PASPT2X, a strictly size-extensive variant. Compared with the intruder-prone PASPT2X, the intruder-free PASPT2H merely neglects second-order corrections that are important only for those states with major projections on the orthogonal complement RX of M0 within the closed space MX ( =M0⊕RX). Such states are usually not supported by the chosen finite one-particle basis set. The weak violation of size-extensivity is therefore numerically insignificant for the target states supported by M0, reinforcing the theoretical basis of PASPT2H.
