Source-linked AI summary
xChemAgents: Agentic AI for Explainable Quantum Chemistry
Can Polat, Mehmet Tuncel, Mustafa Kurban, Erchin Serpedin, Hasan Kurban
TL;DR
Quantum-chemistry surrogates must balance computational efficiency with physical accuracy and interpretable use of heterogeneous chemical metadata. xChemAgents addresses this with Selector–Validator dialogue that sparsifies and validates descriptors before GNN fusion. Experiments show property-dependent gains, with improvements concentrated in electronically governed targets and limited benefit for thermodynamic quantities.
Problem
Accurate first-principles methods are expensive, while geometry-centric GNNs underuse chemical metadata and may lose accuracy or interpretability when descriptors are appended indiscriminately.
Method
xChemAgents uses a Selector to choose weighted descriptors and a Validator to check physical constraints through iterative dialogue before fusing validated features with atomic embeddings.
Results
Electronically governed properties improved consistently with compact string-level descriptors, whereas thermodynamic quantities were largely unaffected or slightly degraded.
Takeaways & Limitations
Late fusion preserves symmetry-aware geometric processing while adding global chemical context, supporting more interpretable multimodal molecular prediction.
Takeaways & Limitations
Evaluation is restricted to QM9, and missing PubChem metadata causes predictions to revert to geometry-only inputs.
Abstract
from arXiv · showhide
Recent progress in multimodal graph neural networks has demonstrated that augmenting atomic XYZ geometries with textual chemical descriptors can enhance predictive accuracy across a range of electronic and thermodynamic properties. However, naively appending large sets of heterogeneous descriptors often degrades performance on tasks sensitive to molecular shape or symmetry, and undermines interpretability. xChemAgents proposes a cooperative agent framework that injects physics-aware reasoning into multimodal property prediction. xChemAgents comprises two language-model-based agents: a Selector, which adaptively identifies a sparse, weighted subset of descriptors relevant to each target, and provides a natural language rationale; and a Validator, which enforces physical constraints such as unit consistency and scaling laws through iterative dialogue. On standard benchmark datasets, xChemAgents achieves up to a 22% reduction in mean absolute error over the state-of-the-art baselines, while producing faithful, human-interpretable explanations. Experiment results highlight the potential of cooperative, self-verifying agents to enhance both accuracy and transparency in foundation-model-driven materials science. The implementation and accompanying dataset are available at https://github.com/KurbanIntelligenceLab/xChemAgents.
1. Introduction
Quantum-chemistry prediction faces a trade-off between accurate but expensive first-principles methods and cheaper approaches with limited transferability. xChemAgents addresses the accuracy, multimodal-feature, and interpretability challenges through cooperative, physics-aware agents.
- DFT provides high-fidelity predictions but its O(N^3) scaling makes systematic studies of large or disordered systems impractical.DFTB reduces cost through tabulated integrals and minimal basis sets, but sacrifices accuracy and transferability outside fitted domains.
- GNN surrogates offer near-DFT accuracy with major speedups, yet commonly ignore textual chemical metadata and can suffer when descriptors are appended naively.Large heterogeneous feature sets may obscure physical relationships and degrade performance on shape- or symmetry-sensitive tasks.
- xChemAgents uses a Selector to choose weighted descriptors and a Validator to check dimensional consistency, scaling laws, and chemical heuristics through up to three dialogue rounds.The validated features are fused with atomic embeddings before GNN prediction, producing an auditable multimodal representation.
- The framework positions interpretability as an agentic reasoning task grounded in domain principles rather than solely as a post-hoc explanation.Its cooperative design builds natural-language rationales and physical validation into descriptor selection.
2. Background
Materials modeling increasingly uses GNN surrogates to accelerate atomistic prediction, but geometry-centric models underuse chemical metadata and remain difficult to interpret. Multi-agent Selector–Validator protocols offer a route to sparse, physically grounded representations.
- GNN-based surrogates use atomic graphs and message passing to capture many-body correlations directly from coordinates, accelerating materials modeling.Representative models include NequIP, TorchMD-Net, QuantumShellNet, and chemically focused CGNN variants.
- Geometry-centric models make limited systematic use of chemical metadata, leaving latent features difficult to relate to bond order, charge transfer, or reaction centers.This opacity hampers error diagnosis, model trust, and adoption by domain scientists.
- Multi-agent systems modularize scientific workflows by assigning complementary roles such as simulation orchestration, convergence tracking, and result triage.Such cooperative protocols have been applied to synthesis planning, catalyst discovery, and battery-electrolyte design.
- The Selector–Validator pattern turns descriptor selection into iterative dialogue, producing sparse, physically grounded, and interpretable representations.A reasoning agent proposes features while a critic checks them against dimensional consistency and known scaling laws.
3. Method
xChemAgents selects a sparse, weighted descriptor subset for each target, validates it against physical criteria through iterative dialogue, and fuses the accepted text representation with an equivariant geometric encoder for prediction.
- 3.1. Agentic Descriptor Selection & Validation: The Selector proposes a target-specific descriptor subset with nonnegative weights and a natural-language rationale, while the Validator returns an acceptance verdict and critique for revision.The descriptor bank contains nine frozen CLIP-embedded descriptors, and the dialogue runs for at most three rounds.
- 3.1. Agentic Descriptor Selection & Validation: Validation checks dimensional consistency, known scaling relations, sparsity, and descriptor complementarity before accepting the representation.The accepted descriptor embeddings are combined by a normalized weighted average, replacing the original CLIP text vector in multimodal fusion.
- 3.2. Equivariant Geometric Encoder: The geometric encoder represents atoms by atomic-number features, connects neighbors within a distance cutoff, and uses message passing with tensor products to preserve rotation equivariance.Angular features use real spherical harmonics to capture three-body geometric interactions, while gated summation aggregates final node embeddings.
- 3.3. Multimodal Fusion and Prediction: A learned gate combines projected geometric and physics-aware text embeddings into a fused vector that a multilayer perceptron maps to the target property.The gate uses concatenation, a sigmoid transformation, and elementwise weighting to determine the contribution of each modality.
- 3.3. Multimodal Fusion and Prediction: Training minimizes mean squared error, while frozen-agent inference adds approximately 0.2 seconds per sample and reduces the descriptor channel from nine to at most five.The reduced channel is intended to improve interpretability and physical alignment without compromising predictive throughput.
4. Experiments
Experiments show that physics-aware multimodal fusion improves electronic-property prediction, while benefits depend on descriptor relevance, fusion placement, and property type.
- 4.1. Effect on Loss: Consistent MAE reductions occur across all backbones for HOMO energy εH, LUMO energy εL, and the HOMO–LUMO gap ∆ε, with smaller improvements for dipole moment µ.Textual annotations provide functional-group and polarity information associated with frontier-orbital behavior and charge separation.
- 4.1. Effect on Loss: Internal energies U0 K and U298 K show limited benefit, with stagnating or increased MAE most visibly for DimeNet++, Equiformer, and FAENet.These properties depend on full potential-energy surfaces and low-frequency vibrational modes, for which high-level descriptors can act as correlated noise.
- 4.1. Effect on Loss: Late fusion in SchNet and DimeNet++ improves five of nine targets, whereas Equiformer’s early fusion benefits orbital tasks but degrades isotropic polarizability α and electronic spatial extent ⟨R2⟩.FAENet shows mixed behavior, improving HOMO energies while reducing accuracy for dipole moments and LUMO levels.
- 4.1. Effect on Loss: The experiments indicate that simple textual descriptors are most effective for electronic observables, while thermodynamic quantities require richer encodings such as force constants or vibrational fingerprints.The appropriate abstraction level of auxiliary information must align with the backbone’s inductive biases and target-property physics.
- 4.2. xChemAgents Decisions: Molecular weight is selected over 20,000 times across targets, but less frequent descriptors receive larger weights for LUMO energy εL and the HOMO–LUMO gap ∆ε.This indicates broad reliance on molecular size and electron count, supplemented by specialized descriptors for refinement.
- 4.2. xChemAgents Decisions: XLogP receives the highest average weight for LUMO energy εL at 0.82, while gross formula receives the highest average weight for the HOMO–LUMO gap ∆ε at 0.64.Polar surface area and hydrogen-bond counts receive moderate weights for polarity-related properties, whereas synonyms are rarely selected and have the lowest weights.
- 4.2. xChemAgents Decisions: For zero-point vibrational energy EZPVE and internal energies U0 K and U298 K, molecular weight dominates selection but all average descriptor weights remain ≤0.35.The Selector distributes weights more evenly across descriptors, consistent with modest or negative MAE changes for these targets.
5. Limitations
The study identifies constraints on xChemAgents’ generality, physical validation, and computational scope.
- QM9 evaluation limits generalization to larger, charged, or inorganic species, while missing PubChem metadata forces geometry-only prediction.
- The Validator primarily checks first-order constraints, leaving quantum interference and stereochemical errors unaddressed.
- Three-round Selector–Validator dialogue adds modest per-molecule latency, and experiments cover only ground-state scalar properties.
6. Discussion & Conclusion
The discussion finds that compact textual descriptors help electronic-property prediction when integrated after geometric processing, but offer limited benefit for thermodynamic targets.
- Electronic-property predictions improved with compact string descriptors, whereas thermodynamic quantities were largely unchanged or slightly degraded.Late fusion worked best in SchNet and DimeNet++, while early fusion diminished accuracy on shape-sensitive targets.
- Feature weights favored chemically informative descriptors, including XLogP for LUMO energy and elemental formula for the HOMO–LUMO gap.These selections highlight polarity and composition as relevant to electronic structure.
- Thermodynamic targets lacked a dominant textual feature, indicating that richer physics-grounded descriptors may be needed for comparable gains.Suggested examples include force constants and vibrational modes.
- xChemAgents demonstrates the potential of cooperative agentic AI for selective, interpretable multimodal chemical representations and future extension beyond ground-state scalars.
Software and Data
The xChemAgents toolkit, descriptors, implementation, and datasets are released under the MIT license using open-access resources.
- The released toolkit includes source code, curated textual descriptors, and the complete Selector–Validator implementation under the MIT license.Datasets and annotations derive exclusively from open-access resources without proprietary restrictions.
Impact Statement
The impact statement presents xChemAgents as a potentially faster and more interpretable framework for virtual screening while identifying risks requiring mitigation.
- Physics-aware language agents integrated with GNNs could support faster, more interpretable virtual screening for green catalyst design and drug discovery.
- Transparent logged rationales support reproducibility and regulatory compliance, while biased descriptors, automated explanations, and training carbon footprint remain risks.The statement proposes balanced datasets, periodic rule audits, and energy-efficient fine-tuning as mitigations.