The presented multimodal transformer networks quantitatively reproduce experimental proton conductivity and the underlying conduction mechanism and provide predictive uncertainty estimates.
The practical utility of a surfactant is determined not only by its equilibrium behavior but also by how quickly it takes effect. Dynamic surface tension (DST)—the time evolution of the surface tension driven by surfactant adsorption onto a freshly formed interface—controls foaming, emulsification, and rapid wetting. Existing techniques for measuring DST are labor-intensive and unsuited for high-throughput screening, hindering surfactant discovery. Here we introduce the first machine-learning framework for predicting DST, providing a valuable complement to experimental approaches. Trained on a purpose-built, curated database of 979 DST series for 233 surfactants, the domain-specific chemical language model reproduces DST with a mean absolute error of 6.5 mN m –1 and predicts the characteristic half-time of the surface-tension relaxation with a coefficient of determination of 0.78. A hybrid aleatoric–epistemic uncertainty quantification explicitly delineates the applicability domain and, by discarding high-variance predictions, lowers the mean absolute error to 5.6 mN m –1 . Finally, a game-theoretic (Shapley) attribution shows that the model autonomously recovers fragment-level rules that are consistent with colloid-chemical knowledge. Our framework, combining physics-informed training, rigorous uncertainty estimation, and game-theoretic interpretability, offers a general strategy for tackling time-dependent interfacial phenomena. The curated database, the trained model, and an interactive online predictor are released publicly.
The accurate simulation of photochemical reactions requires methods that capture nonadiabatic transitions through conical intersections between different electronic states. While machine-learning interatomic potentials (MLIPs) offer a promising route to efficient nonadiabatic molecular dynamics, their training for excited states is often limited by the prohibitive cost of generating extensive datasets at a sufficiently high level of quantum chemistry theory. Here, we explore strategies for developing MLIPs that achieve multistate multireference perturbation theory accuracy, systematically comparing single-state, multi-output, multi-state, transfer learning, and delta-learning architectures. We show that transfer learning (TL) from CASSCF to XMCQDPT2 provides the best balance of accuracy and computational efficiency, dramatically reducing the amount of expensive reference data required. This methodology is validated on the methaniminium cation, whose complete photodissociation landscape, including S1 branching into photoisomerization and a direct H2-loss pathway mediated by a recently discovered conical intersection, is captured at the XMCQDPT2/SA(3)-CASSCF(12,12) level. We show that final product yields are largely independent of training strategy, whereas the TL model produces distinct ultrafast population dynamics compared to the randomly initialized model. Finally, we develop a wavepacket oscillation model for fitting excited-state population dynamics, which quantitatively reproduces the ultrafast non-exponential decay and extracts state- and channel-specific lifetimes, directly linking quantum transition probabilities to classical rate constants.