Mumin Enis Leblebici | Searcharxiv

arXiv2026

MultiPUFFIN: A Multimodal Domain-Constrained Foundation Model for Molecular Property Prediction of Small Molecules

MultiPUFFIN is a domain-informed multimodal foundation model for predicting thermophysical properties of small molecules, addressing a critical gap in chemical engineering, drug discovery, and materials science. Existing molecular foundation models pretrain on millions of molecules to learn general-purpose representations, but their standard MLP output layers impose no physical constraints, vapor pressure predictions may violate monotonic temperature dependence, and viscosity curves may lack the functional form required by process simulators. Domain-informed approaches that guarantee thermodynamic consistency have remained limited to single properties and small datasets, whereas multimodal foundation models have focused on biological activity rather than thermophysical properties. MultiPUFFIN fills this gap by fusing SMILES sequences, 2D molecular graphs, and 3D conformer geometries through bidirectional cross-modal attention and gated fusion, supplemented by auxiliary encoders for experimental conditions and molecular descriptors. The backbone is pretrained on 500,000 unlabelled PubChem molecules using three complementary self-supervised objectives. A condition-aware refinement stack of five conditioners (temperature, pH, pressure, polymorph, and measurement method) routes each property to a four-head tournament that selects the best-performing thermodynamically informed head for that property. MultiPUFFIN achieves a mean test R2 of 0.784 and outperforms fine-tuned ChemBERTa-2 on all nine properties despite training on roughly 2,000x fewer labeled molecules.

cs.LG↗

arXiv2025

ExPUFFIN: Thermodynamic Consistent Viscosity Prediction in an Extended Path-Unifying Feed-Forward Interfaced Network

Accurate prediction of liquid viscosity is essential for process design and simulation, yet remains challenging for novel molecules. Conventional group-contribution models struggle with isomer discrimination, large molecules, and parameter availability, while purely data-driven graph neural networks (GNNs) demand large datasets and offer limited interpretability. Even when feasible to be applied, purely data-driven models lack thermodynamic consistency in their predictions and are not a reliable solution. This work introduces ExPUFFIN, an extended version of the Path-unifying Feed-Forward Interfaced Network, consisting of a hybrid GNN-based framework that directly predicts temperature-dependent viscosities of pure hydrocarbons from molecular graphs, while enforcing mechanistic inductive biases in the output layer to ensure thermodynamic consistency. Molecular information is given as graph structures, encoded as a graph convolutional network, and mapped to an inductive bias neuron based on two thermophysical correlations: a three-parameter Andrade-type equation and a four-parameter empirical viscosity-temperature relation. The accuracy of these models is compared with a solely data-driven prediction. The Andrade-based ExPUFFIN variant reduces RMSE compared to the purely data-driven baseline of 37 percent and yields smooth, physically consistent interpolation and extrapolation of viscosity-temperature curves, properties that are not observed in purely data-driven models. The empirical ExPUFFIN model provides comparable accuracy while retaining robust trends. Overall, embedding physics-based structure in GNN outputs improves accuracy, robustness, and transferability, enabling reliable viscosity predictions for complex hydrocarbon molecules. The approach is readily extendable to other properties and significantly broader chemical domains.

physics.chem-ph↗