Abstract
Predicting physicochemical properties of small molecules is central to chemical engineering, drug discovery and materials science. Existing molecular foundation models pretrain on millions of molecules, but their standard output layers impose no physical constraints, so predicted vapour-pressure or viscosity curves can violate the functional form process simulators require; domain-informed models that guarantee consistency have been limited to single properties, and multimodal foundation models target biological activity rather than thermophysical properties. We introduce MultiPUFFIN, a domain-informed multimodal foundation model that fuses SMILES sequences, 2D graphs and 3D conformers through bidirectional cross-modal attention and gated fusion, with auxiliary encoders for experimental conditions and molecular descriptors. The backbone is pretrained on 500,000 unlabelled PubChem molecules by self-supervision; a condition-aware refinement stack routes each property to a four-head tournament that selects the best thermodynamically-informed head, and cross-property couplings enforce inter-property consistency. On a scaffold-split test set of 8877 molecules from 37,968 unique molecules across nine thermophysical properties, MultiPUFFIN reaches an in-scope mean test š¯‘…2 = 0.784. On the six temperature-independent properties, where the comparison with fine tuned ChemBERTa-2 is symmetric, it improves mean š¯‘…2 from 0.658 to 0.769 while using roughly 2000Ć— fewer labelled molecules; the temperaturedependent properties are reported separately. Every reported metric carries a bootstrap confidence interval, and per-property applicability-domain filters bound the deployment scope. Beyond accuracy, the equation heads make monotonic temperature dependence and admissible coefficient signs hold by construction: permuting the equationā€“property assignment degrades vapor-pressure accuracy by 42%, confirming the constraints are loadbearing.