betaFeedback

WorksComputer Science

xChemAgents: Agentic AI for Explainable Quantum Chemistry

Abstract

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.

Introduction

First-principle electronic structure methods such as density functional theory (DFT) remain the gold standard for predicting bonding, defect energetics, and phase stability. However, the cubic wall-time scaling O(N3)\mathcal{O}(N^3) with electron count due to repeated Hamiltonian diagonalization makes systematic studies of large or disordered systems impractical [8] (Cohen et al. 2012). Density functional tight binding (DFTB) offers a more tractable alternative by tabulating two-center integrals and using minimal basis sets, but this reduction in cost comes at the expense of accuracy and transferability. Benchmark studies reveal large errors for long-range dispersion, many-body polarization, and strongly correlated states outside the fitted parameter domain (Hourahine et al., 2020; Goyal et al., 2014). These trade-offs sustain the search for surrogate models that preserve quantum-mechanical fidelity while scaling to diverse, chemically rich systems.

Graph neural networks (GNNs) have emerged as promising surrogates by operating directly on atomic graphs or periodic neighbor lists, achieving near-DFT accuracy with orders-of-magnitude speedups [16]. Yet most state-of-the-art GNNs remain unimodal, focusing solely on Cartesian geometry while ignoring rich textual metadata—empirical descriptors, spectral signatures, and systematic names—available in curated databases like PubChem [1]. Naïvely appending such descriptors can inflate the feature space, obscure physical relationships, and even degrade performance on symmetry-sensitive or long-range electrostatic tasks [13]. These challenges exacerbate an ongoing interpretability deficit in materials machine learning, where node-level saliency maps rarely map cleanly to chemically meaningful insights [9].

In parallel, recent advances in natural language processing have demonstrated that agentic AI—systems composed of interacting large language model (LLM) agents—can solve complex reasoning tasks through cooperation, self-critique, and dialog-based refinement [22] (Dorri et al. 2018; Li et al. 2024). Multi-agent orchestration enables autonomous scientific workflows for experiment planning, data analysis, and hypothesis generation, while reducing hallucination and improving robustness through inter-agent feedback loops [24]. These developments reframe feature selection for scientific surrogates not just as an optimization problem, but as an opportunity to embed domain knowledge and interpretability into the learning pipeline via deliberative, agentic reasoning.

This work presents xChemAgents, a cooperative multi-agent framework that embeds physics deliberation into multimodal molecular representation learning. A Selector agent, implemented as a chemistry-tuned language model, queries a descriptor pool and proposes a compact, weighted subset of features for each target property, accompanied by a natural language rationale. A complementary Validator agent enforces physical constraints—dimensional consistency, scaling laws, and chemical heuristics—and either approves the proposal or returns structured feedback for revision. This dialogue continues for up to three rounds, after which the validated features are fused with atomic embeddings and passed to a GNN surrogate. The result is a hybrid input that is not only more accurate, but also physically interpretable and auditable. The overall process is summarized in Figure 1. xChemAgents exemplifies a shift toward agentic AI in scientific domains—transforming interpretability from a post-hoc explanation challenge into a multi-agent reasoning task grounded in domain principles.

Flowchart of the Selector–Validator pipeline for molecular feature selection

Figure 1. Flowchart of the Selector–Validator pipeline for molecular feature selection. The process begins with an input molecule and its property description; the Selector agent identifies 3–5 candidate features, assigns normalized weights, and documents its reasoning. The Validator agent then evaluates feature relevance, weight accuracy, and overall completeness. If validation succeeds, the finalized, validated selections are output; if validation fails, critique feedback is returned to the Selector for iterative refinement.

The remainder of the paper is organized as follows. § 2 reviews background material. § 3 describes the method. § 4 presents experiments. § 5 discusses limitations. § 6 concludes the paper.

Background

Ab-Initio Simulations

DFT maps the interacting many-electron problem onto a set of non-interacting single-particle equations solved self-consistently [3]. Each self-consistency iteration requires diagonalizing the Kohn–Sham Hamiltonian—a “wall” that grows with the number of basis functions [5]. While sparse-matrix techniques and linear-scaling (O(N)\mathcal{O}(N)) variants exploit locality to reduce formal complexity for large-gap (insulating) systems, metallic or highly disordered materials still demand dense algebra, so conventional DFT codes routinely incur wall-times of hours to days once system sizes exceed a few hundred atoms (Hernández et al., 1996). Approximations such as DFTB perform a Taylor expansion of the total DFT energy around a reference density and tabulate two-centre integrals to achieve one- to two- orders of magnitude speed-ups over full DFT, but they inherit systematic errors—most notably missing long-range dispersion, incomplete many-body polarization, and challenges in treating strongly correlated states—without empirical corrections or higher-order expansions (Gaus et al., 2011).

Machine Learning for Materials Science

Machine learning surrogates have become the workhorse accelerators of contemporary materials modelling. Interatomic potentials and property predictors constructed with GNNs treat atoms as graph nodes and bonded or spatial neighbours as edges, enabling message-passing operations that capture complex many-body correlations directly from atomic coordinates. Models such as NequIP (Batzner et al., 2022), TorchMD-Net [21], QuantumShellNet [12], and chemically focused CGNN variants (Cheng et al., 2021) demonstrated that deep geometric learning can reproduce density-functional energies and forces at millisecond speed. Subsequent advances in equivariant design—exemplified by E(n)-GNN [18], SE(3)-Transformer (Fuchs et al., 2020), PaiNN [19], SphereNet (Coors et al., 2018), and GotenNet (Aykent & Xia, 2025)—embed rotational and permutational symmetry directly into the network, dramatically improving data efficiency and stability. More recently, multimodal encoders such as Pure2DopeNet [11] and CrysMMNet (Das et al., 2023) fuse structural graphs with textual and spectral channels, while generative pipelines like CDVAE [25] and CrystalFlow [6] learn to sample chemically valid molecules or crystals under explicit geometric constraints. Hierarchical pooling and attention mechanisms extend the effective receptive field to hundreds of atoms, yet—even with these innovations—the overwhelming majority of models remain geometry centric where they make little systematic use of the rich chemical metadata available in public repositories, and their latent features are difficult to map onto intuitive notions such as bond order, charge transfer or reaction centre. This opacity hampers error diagnosis, model trust and, ultimately, adoption by domain scientists.

Multi-Agent Systems

Multi-agent systems provide a principled framework for modularizing complex scientific tasks [7]. Recent work has demonstrated the integration of agents with domain-specific tools—such as one agent orchestrating simulations, another tracking convergence criteria, and a third triaging results and suggesting follow-up experiments [20]. Cooperative protocols have been successfully applied in diverse scientific domains, including autonomous synthesis planning, closed-loop catalyst discovery, and the inverse design of battery electrolytes (Jia et al., 2025).

Within representation learning, a prominent Selector–Validator pattern has emerged: a reasoning agent proposes candidate features or hypotheses, while a critic agent validates them against physical constraints such as dimensional consistency or known scaling laws. The xChemAgents framework exemplifies this paradigm by transforming descriptor selection from a static preprocessing step into an iterative, agent-driven dialogue. This leads to sparse, physically grounded, and interpretable representations that serve as inputs to downstream GNN-based surrogate models [26].

Method

Agentic Descriptor Selection & Validation

Let D={ϕ1,…,ϕm}\mathcal{D} = \{\phi_1, \ldots, \phi_m\} denote the full bank of textual descriptors provided with the enriched QM9 dataset [17, 15], where m=9m = 9 in this study. Each descriptor ϕk\phi_k is embedded once using a frozen CLIP encoder [14], yielding a fixed vector representation ϕk∈Rdϕ\phi_k \in\mathbb{R}^{d_\phi}.

Selector Agent. Given an input molecule xx and a target property yy, the Selector agent is a function

Asel:(x,y)↦(S,w,rsel),A_{\mathrm{sel}} : (x, y) \mapsto(S, \mathbf{w}, r_{\mathrm{sel}}),

which outputs:

  1. a rationale rselr_{\mathrm{sel}} in natural language (chain-of-thought),

  2. a selected subset S={ϕk1,…,ϕkp}⊂DS = \{\phi_{k_1}, \ldots, \phi_{k_p}\} \subset\mathcal{D} with p∈{3,4,5}p \in\{3, 4, 5\},

  3. a weight vector w∈Δp−1\mathbf{w} \in\Delta^{p-1}, where Δp−1={w∈Rp∣∑i=1pwi=1,  wi≥0}\Delta^{p-1} = \{\mathbf{w} \in\mathbb{R}^{p} \mid\sum_{i=1}^{p} w_i = 1,\; w_i \ge0\} is the (p−1)(p-1)-dimensional probability simplex.

Validator Agent

Validator Agent. The Validator agent

Aval:(S,w,y)↦(v,rval)A_{\mathrm{val}} : (S, \mathbf{w}, y) \mapsto(v, r_{\mathrm{val}})

accepts the selected descriptors SS, their weights w\mathbf{w}, and the target property yy, returning:

  • a binary verdict v∈{0,1}v \in\{0,1\} indicating acceptance (1) or rejection (0),

  • a critique rvalr_{\mathrm{val}} in natural language explaining the reasoning behind vv.

The Validator is guided by structured prompts encoding:

  1. unit consistency between each ϕk\phi_k and yy,

  2. adherence to known scaling relations (e.g., Koopmans’ theorem for HOMO/LUMO predictions),

  3. sparsity and complementarity heuristics to prevent redundancy and encourage diversity among descriptors.

If v=0v = 0, the critique rvalr_{\mathrm{val}} is appended to the dialogue history, and the Selector re-samples. This dialogue loop runs for at most L=3L = 3 rounds (see Algorithm 1).

Physics-aware textual embedding. Given the accepted descriptor set S^={ϕk1,…,ϕkp}\hat{S} = \{\phi_{k_1}, \ldots, \phi_{k_p}\} and corresponding normalized weights w^\hat{\mathbf{w}}, the final physics-aware text embedding is computed as a weighted average:

tphys=∑j=1pw^jϕkj,(1)\mathbf{t}_{\mathrm{phys}} = \sum_{j=1}^{p} \hat{w}_j \phi_{k_j}, \tag*{(1)}

where ϕkj∈Rdϕ\phi_{k_j} \in\mathbb{R}^{d_\phi} denotes the frozen CLIP embedding of descriptor ϕkj\phi_{k_j}. This representation replaces the original CLIP text vector in the multimodal fusion module (§ 3.4), ensuring that only physics-vetted information influences the downstream prediction.

Equivariant Geometric Encoder

Each atom is represented as a node ii with an initial feature vector hi(0)=onehot⁡(Zi)\mathbf{h}_i^{(0)} = \operatorname{onehot}(Z_i), where ZiZ_i is the atomic number. Edges (i,j)(i,j) connect all atomic pairs within a cutoff radius rcr_c, based on interatomic distances dij=∥ri−rj∥d_{ij} = \lVert\mathbf{r}_i - \mathbf{r}_j \rVert.

Distances are expanded using Bessel radial basis functions (RBFs):

ρ(dij)=[ρ1(dij),…,ρB(dij)],\rho(d_{ij}) = [\rho_1(d_{ij}), \ldots, \rho_B(d_{ij})],

Algorithm 1 xChemAgents: End-to-End Physics-Vetted Inference

  1. Input: atomic positions {ri}\{r_i\}, atomic numbers {Zi}\{Z_i\}, target property yy

  2. Require: descriptor bank D\mathcal{D}, selector agent Asel\mathcal{A}_{\mathrm{sel}}, validator Aval\mathcal{A}_{\mathrm{val}}, equivariant GNN Gθ\mathcal{G}_{\theta}, fusion head Fθ\mathcal{F}_{\theta}

  3. Output: prediction y^\hat{y}, accepted feature set S^\hat{S}, weights w^\hat{w}, rationales

  4. for ℓ=1\ell= 1 to LL do

  5. (S,w,rsel)←Asel(y,D)(S,w,r_{\mathrm{sel}}) \leftarrow\mathcal{A}_{\mathrm{sel}}(y,\mathcal{D})

  6. (v,rval)←Aval(S,w,y)(v,r_{\mathrm{val}}) \leftarrow\mathcal{A}_{\mathrm{val}}(S,w,y)

  7. if vv then

  8. S^←S, w^←w\hat{S} \leftarrow S,\ \hat{w} \leftarrow w

  9. break

  10. end if

  11. end for

  12. tphys←∑ϕk∈S^w^kCLIP⁡(ϕk)t_{\mathrm{phys}} \leftarrow\sum_{\phi_k \in\hat{S}} \hat{w}_k \operatorname{CLIP}(\phi_k)

  13. initialise hi(0)←onehot⁡(Zi)h_i^{(0)} \leftarrow\operatorname{onehot}(Z_i)

  14. for t=0t = 0 to T−1T-1 do

  15. for all edges (i,j)(i,j) with dij<rcd_{ij} < r_c do

  16. mij(t)←TP⁡(hi(t),hj(t),RBF⁡(dij))m_{ij}^{(t)} \leftarrow\operatorname{TP}(h_i^{(t)},h_j^{(t)},\operatorname{RBF}(d_{ij}))

  17. end for

  18. hi(t+1)←hi(t)+∑j∈N(i)mij(t)h_i^{(t+1)} \leftarrow h_i^{(t)} + \sum_{j \in\mathcal{N}(i)}m_{ij}^{(t)}

  19. end for

  20. g←∑iσ(Wgatehi(T))⊙hi(T)g \leftarrow\sum_i \sigma(W_{\mathrm{gate}}h_i^{(T)}) \odot h_i^{(T)}

  21. t~←WtPtphys;g~←Wgg\tilde{t} \leftarrow W_tPt_{\mathrm{phys}};\quad\tilde{g} \leftarrow W_gg

  22. normalise (g~,t~)(\tilde{g},\tilde{t})

  23. z←σ(W[g~∥t~]+b)z \leftarrow\sigma(W[\tilde{g}\Vert\tilde{t}]+b)

  24. f←z⊙g~+(1−z)⊙t~f \leftarrow z \odot\tilde{g}+(1-z)\odot\tilde{t}

  25. y^←Fθ(f)\hat{y} \leftarrow\mathcal{F}_{\theta}(f)

where ρb(⋅)\rho_b(\cdot) denotes the bb-th basis function. Angular features θijk\theta_{ijk} are similarly embedded using real spherical harmonics Yℓm(θijk)Y_{\ell m}(\theta_{ijk}) to capture 3-body geometric interactions.

At message-passing step t∈{0,…,T−1}t \in\{0,\ldots,T-1\}, node embeddings are updated as:

mij(t)=TP⁡(hi(t),hj(t),ρ(dij))(2)m_{ij}^{(t)}=\operatorname{TP}(h_i^{(t)},h_j^{(t)},\rho(d_{ij})) \tag*{(2)}
hi(t+1)=hi(t)+∑j∈N(i)mij(t)(3)h_i^{(t+1)}=h_i^{(t)}+\sum_{j\in\mathcal{N}(i)}m_{ij}^{(t)} \tag*{(3)}

where TP denotes a learnable tensor product that mixes irreducible representations of SO(3), ensuring equivariance under global rotations.

After TT message-passing layers, the final node embeddings are aggregated via gated summation:

g=∑iσ(Wgatehi(T))⊙hi(T)∈Rn(4)g=\sum_i\sigma(W_{\mathrm{gate}}h_i^{(T)})\odot h_i^{(T)}\in\mathbb{R}^{n} \tag*{(4)}

where Wgate∈Rn×dW_{\mathrm{gate}}\in\mathbb{R}^{n\times d} is a learnable projection matrix, σ(⋅)\sigma(\cdot) denotes the sigmoid activation, and ⊙\odot indicates elementwise multiplication.

Multimodal Fusion and Prediction

The text-derived embedding tphyst_{\mathrm{phys}} (see Eq. (1)) is projected into a dd-dimensional latent space:

tp=Ptphys,P∈Rd×768.(5)t_p=Pt_{\mathrm{phys}},\quad P\in\mathbb{R}^{d\times768}. \tag*{(5)}

The geometric embedding gg (from Eq. (4)) is projected using learned linear transformations:

gproj=Wgg,tproj=Wttp,(6)g_{\mathrm{proj}}=W_gg,\quad t_{\mathrm{proj}}=W_tt_p, \tag*{(6)}

where Wg,Wt∈Rd×dW_g,W_t\in\mathbb{R}^{d\times d}. Both embeddings are normalized using layer normalization:

g~=LN⁡(gproj),t~=LN⁡(tproj).\tilde{g}=\operatorname{LN}(g_{\mathrm{proj}}),\quad\tilde{t}=\operatorname{LN}(t_{\mathrm{proj}}).

A gating mechanism computes elementwise fusion weights:

z=σ(W[g~∥t~]+b),z=\sigma\left(W[\tilde{g}\Vert\tilde{t}]+b\right),

where W∈Rd×2dW\in\mathbb{R}^{d\times2d} is a learned matrix, b∈Rdb\in\mathbb{R}^{d} is a bias vector, ∥\Vert denotes concatenation, and σ(⋅)\sigma(\cdot) is the sigmoid function.

The fused representation is computed as:

f=z⊙g~+(1−z)⊙t~,(7)f=z\odot\tilde{g}+(1-z)\odot\tilde{t}, \tag*{(7)}

where ⊙\odot denotes elementwise multiplication. This vector f∈Rdf\in\mathbb{R}^{d} is passed to a multi-layer perceptron fθf_{\theta} to predict the target property:

y^=fθ(f).(8)\hat{y}=f_{\theta}(f). \tag*{(8)}

The model is trained to minimize the mean squared error between y^\hat{y} and the ground-truth property. During training and inference, both agents remain frozen. The agent dialogue process (Algorithm 1) introduces an average overhead of approximately 0.2 seconds per sample but reduces the descriptor channel from 9 to a maximum of 5. This enables improved interpretability and physical alignment without compromising predictive throughput.

Experiments

Dataset. The QM9 benchmark consists of approximately 134k small organic molecules stored as XYZ files, which capture only their three-dimensional geometries. To provide the chemical context absent from atomic coordinates, each molecule is augmented with metadata retrieved from PubChem, including IUPAC names, molecular formulas and masses, XLogP, counts of hydrogen-bond donors and acceptors, rotatable bond counts, topological polar surface area, formal charge, synonyms, and selected spectral annotations. This results in a multimodal dataset that integrates geometric and textual descriptors to enable richer property prediction. Molecules lacking any of the required PubChem fields are excluded from the final corpus.

Implementations

Evaluation is conducted on the QM9 benchmark using four backbone architectures that span diverse regions of the contemporary GNN design space: (i) SchNet; (ii) DimeNet++, which employs directional message passing and spherical harmonics; (iii) Equiformer, an E(3)-equivariant graph-attention transformer; and (iv) FAENet, a frame-averaging network without explicit SO(3) constraints. To ensure architectural consistency, all backbones are equipped with a shared gated-fusion block, isolating the impact of the agentic feature selector from variations in model capacity. Networks are implemented using PyTorch Paszke et al., [10], PyTorch Geometric Fey & Lenssen, 2019, and the official repositories. Optimization is performed using the ADAM optimizer Kingma & Ba, [2] with an ℓ1\ell_1 loss, a learning rate of 10−310^{-3}, a batch size of 64, and 35 training epochs. Performance is reported as the mean over three randomized folds (k=3k = 3) to account for dataset variability. The agentic system is implemented using LangChain, with Ollama serving as the backbone model and CLIP as the tokenizer, integrated via the Transformers library Wolf et al., [23].

Gated-fusion layer. Each backbone outputs a graph-level embedding g∈R128\mathbf{g} \in\mathbb{R}^{128}, which is fused with a physics-vetted CLIP embedding tphys∈R32\mathbf{t}_{\mathrm{phys}} \in\mathbb{R}^{32}. The fusion is performed via a learnable gating mechanism:

f=σ(W[g∥tphys])⊙Wgg+(1−σ(⋅))⊙Wttphys,\mathbf{f} = \sigma\left(\mathbf{W}[\mathbf{g}\Vert\mathbf{t}_{\mathrm{phys}}]\right) \odot\mathbf{W}_{g}\mathbf{g} + \left(1-\sigma(\cdot)\right) \odot\mathbf{W}_{t}\mathbf{t}_{\mathrm{phys}},

where [⋅∥⋅][\cdot\Vert\cdot] denotes concatenation, σ(⋅)\sigma(\cdot) is the sigmoid function, and ⊙\odot denotes element-wise multiplication. The resulting multimodal vector f\mathbf{f} is passed to a two-layer MLP for final property prediction.

SchNet. The SchNet configuration uses six interaction blocks, a hidden size of 128, 50 Gaussian filters, and a cutoff radius of 10 Å. Global add-pooling is applied to produce a 128-dimensional graph-level embedding, which is fused with the CLIP vector via the gated-fusion layer. A single linear layer outputs the scalar prediction.

DimeNet++. DimeNet++ is configured with hyperparameters scaled down by half relative to the original implementation: three interaction blocks, 128 hidden channels, a spherical rank of lmax⁡=7l_{\max} = 7, radial basis size of 6, and a cutoff radius of 5 Å. The default readout is replaced with the gated-fusion layer followed by a linear prediction head.

Equiformer. The Equiformer employs four attention layers with two heads and uses 50% fewer representation channels than the original model, resulting in approximately 4.8M parameters—comparable to FAENet. The graph-level embedding has a dimensionality of 256, which is fused with a 32-dimensional CLIP branch and passed through a linear layer for prediction.

FAENet. FAENet is configured with a cutoff radius of 6 Å, two interaction blocks, a hidden dimension of 128, and 20 Gaussian distance channels. The textual CLIP embedding (768-dimensional) is reduced to 32 dimensions via a two-layer projection head before being fused with the graph representation.

Effect on Loss

The inclusion of textual annotations—including IUPAC names, chemical formulas, molecular weights, rotatable bond counts, polar surface areas, hydrogen-bond donor/acceptor counts, and synonyms—introduces high-level information related to functional groups and molecular polarity. These descriptors correlate strongly with frontier orbital behavior, resulting in consistent reductions in mean absolute error (MAE) across all backbones for HOMO energy εH\varepsilon_{\mathrm{H}}, LUMO energy εL\varepsilon_{\mathrm{L}}, and the HOMO–LUMO gap Δε\Delta\varepsilon, as shown in Table 1. Smaller but noticeable improvements are also observed for the dipole moment μ\mu, whose magnitude reflects charge separation patterns captured by the same descriptors.

Table 1. Mean Absolute Error (MAE; lower is better) for molecular property prediction across baseline architectures and their xChemAgents-enhanced variants. Models augmented with xChemAgents are highlighted in bold, reflecting consistent performance improvements through physics-aware cooperative reasoning.

TargetSchNet BaseSchNet xChemAgentsDimeNet++ BaseDimeNet++ xChemAgentsEquiformer BaseEquiformer xChemAgentsFAENet BaseFAENet xChemAgents
μ\mu (D)0.23080.21390.17470.16400.11540.12450.26500.2820
α\alpha (Å3^{3})0.31540.41940.37620.34310.28870.36060.49250.4675
εH\varepsilon_{\mathrm{H}} (eV)0.17900.13580.13170.11250.13800.11260.14240.1377
εL\varepsilon_{\mathrm{L}} (eV)0.13120.10270.08910.08040.08140.07630.09800.1014
Δε\Delta\varepsilon (eV)0.22710.17790.17260.14890.16270.14570.19320.1928
⟨R2⟩\langle R^{2}\rangle (bohr2^{2})5.47416.24219.64629.08995.59836.655915.006813.4011
EZPVEE_{\mathrm{ZPVE}} (eV)0.01150.01060.00670.00710.00940.00960.00950.0107
U0 KU_{0\ \mathrm{K}} (eV)30.958228.328339.706842.102512.391319.55729.806521.2594
U298 KU_{298\ \mathrm{K}} (eV)28.150732.654132.748641.309913.038826.767711.020419.8619

Table 1.

In contrast, internal energies U0KU_{0\mathrm{K}} and U298KU_{298\mathrm{K}}, which depend on full potential-energy surfaces and low-frequency vibrational modes, show limited benefit. Here, the high-level descriptors contribute little and often act as correlated noise, resulting in stagnating or increased MAE—most visibly for DimeNet++, Equiformer, and FAENet.

The location and frequency of fusion modulate these trends.

SchNet and DimeNet++ apply the annotation vector only at the global readout stage, allowing auxiliary information to influence the output without disrupting symmetry-aware message passing. This late-fusion strategy leads to net improvements on five of nine targets. Equiformer, by contrast, injects the annotation vector at every attention layer, repeatedly mixing non-geometric tokens into equivariant updates. This early-fusion approach benefits orbital-related tasks but degrades geometry-dominated properties such as isotropic polarizability α\alpha and electronic spatial extent ⟨R2⟩\langle R^{2}\rangle. FAENet adopts an intermediate strategy and exhibits mixed behavior—improving predictions for HOMO energies while reducing accuracy for dipole moments and LUMO levels.

These results suggest that simple text-based descriptors are most effective for properties governed by electronic distribution, whereas thermodynamic quantities demand richer, physics-aware encodings such as force constants or vibrational fingerprints. The findings also underscore the importance of aligning the abstraction level of auxiliary information with the backbone’s inductive biases: the same cue that improves one prediction can degrade another if integrated incongruently. A percentile-based comparison is provided in Figure 2.

Figure 2. Percentile change in MAE across models for each target property. Positive values denote an increase in error relative to the baseline, while negative values indicate an improvement in the model’s performance.

xChemAgents Decisions

Table 2 summarizes the behavior of the Selector agent across all tasks. The first row records the frequency with which each textual descriptor is selected, while the second row reports the mean normalized weight assigned when the feature is included.

TargetMetricIUPACFormulaMolecular WeightXLogPH-Bond DonorsH-Bond AcceptorsRotatable BondsPSASynonyms
μ\muSelection Count23402734215971337814935168938134<u>18146</u>79
Normalized Importance0.260.26<u>0.28</u>0.240.220.240.230.290.18
α\alphaSelection Count9512328222231457664589241<u>18344</u>1581681
Normalized Importance0.260.270.310.260.200.210.25<u>0.28</u>0.18
εH\varepsilon_{\mathrm{H}}Selection Count266135482120515724994210478<u>16430</u>11820190
Normalized Importance0.270.270.34<u>0.31</u>0.230.230.300.270.21
εL\varepsilon_{\mathrm{L}}Selection Count218531941906816700968013294<u>17484</u>10650175
Normalized Importance0.260.26<u>0.28</u>0.820.220.230.250.260.20
Δε\Delta\varepsilonSelection Count14552537204971642585709592<u>17726</u>10453157
Normalized Importance0.260.64<u>0.31</u>0.260.230.220.280.280.18
⟨R2⟩\langle R^2\rangleSelection Count100115442206475533022599221634<u>21964</u>117
Normalized Importance0.270.280.390.250.210.220.28<u>0.29</u>0.19
EZPVEE_{\mathrm{ZPVE}}Selection Count111525942309294771024910335<u>16997</u>15379113
Normalized Importance0.260.280.350.240.270.250.26<u>0.31</u>0.18
U0KU_{0\mathrm{K}}Selection Count1116266822606117691056310494<u>15230</u>1198083
Normalized Importance0.260.440.350.290.230.230.280.270.17
U298KU_{298\mathrm{K}}Selection Count72816142392514419120169394<u>15722</u>1307374
Normalized Importance0.25<u>0.28</u>0.330.260.220.230.240.270.18

Table 2. Summary of descriptor relevance across target properties. Row 1 reports the number of times each descriptor was selected by the Selector agent across runs. Row 2 shows the average normalized importance weights assigned to each descriptor. Best values in each row are shown in bold, and second-best are <u>underlined</u>.

A property-agnostic trend emerges: molecular weight overwhelmingly dominates the selection statistics, being chosen over 20,000 times across all electronic and thermodynamic targets—far more than any other descriptor. This reflects the agent’s reliance on a coarse proxy for molecular size and electron count in the absence of more specific physical signals.

However, the normalized weights reveal a more nuanced behavior. High selection frequency does not necessarily imply high influence. For several tasks—most notably LUMO energy εL\varepsilon_{\mathrm{L}} and the HOMO–LUMO gap Δε\Delta\varepsilon—less frequently chosen descriptors are assigned significantly larger weights when selected. This suggests that the agent defaults to molecular weight for general coverage but relies on more specialized descriptors to refine predictions where they matter most.

Electronic descriptors. For LUMO energy εL\varepsilon_{\mathrm{L}}, the lipophilicity index XLogP receives the highest average weight (0.82), nearly three times greater than any other descriptor, despite being selected only 16,700 times compared to 19,068 selections for molecular weight. A high XLogP typically indicates the presence of extended hydrophobic and often conjugated fragments, which stabilize low-lying unoccupied orbitals—explaining its elevated influence on εL\varepsilon_{\mathrm{L}}.

For the HOMO–LUMO gap Δε\Delta\varepsilon, a complementary trend emerges: the gross formula achieves the highest mean weight (0.64), suggesting that elemental composition—especially the inclusion of heteroatoms—serves as a dominant discriminator once molecular size is accounted for.

Properties such as dipole moment μ\mu and isotropic polarizability α\alpha depend on broader polarity cues. Here, polar surface area and hydrogen-bond donor/acceptor counts are assigned moderate but consistent weights, reflecting their connection to charge separation and electron cloud deformability. In contrast, Synonyms, being purely linguistic metadata, are rarely selected and consistently assigned the lowest weights, underscoring their limited physical relevance.

Thermodynamic quantities. Zero-point vibrational energy EZPVEE_{\mathrm{ZPVE}} and internal energies U0 KU_{0\ \mathrm{K}} and U298 KU_{298\ \mathrm{K}} exhibit a strong selection bias toward molecular weight, consistent with the scaling of vibrational degrees of freedom with system size. However, their average assigned weights (all ≤0.35\le0.35) remain notably lower than those observed for frontier-orbital tasks, reinforcing the earlier observation that high-level string descriptors offer limited signal for properties governed by detailed potential-energy surfaces.

For these targets, the Selector agent distributes weights more evenly across descriptors such as XLogP, hydrogen-bond statistics, and polar surface area, indicating that no single textual feature provides dominant predictive power. This behavior suggests that the agent is effectively “hedging” by forming diluted ensembles—a strategy consistent with the modest or negative MAE changes observed in Table 1.

Overall, the feature selection patterns support the conclusion that simple text-based descriptors are most effective for electronic observables. In contrast, thermodynamic properties appear to require richer, physics-aware representations—such as force constants or vibrational fingerprints—to achieve comparable predictive improvements.

Limitations

xChemAgents is currently evaluated only within the constrained chemical space of QM9, which limits generalization to larger, charged, or inorganic species. The system also relies on the availability and quality of PubChem descriptors; in cases where these metadata are missing, predictions revert to geometry-only inputs.

The Validator agent enforces primarily first-order physical constraints, such as unit consistency and selected scaling laws. More nuanced issues—such as quantum interference effects or stereochemical errors—are not explicitly addressed.

Additionally, the three-round Selector–Validator dialogue introduces modest computational overhead, adding latency per molecule. This study focuses exclusively on ground-state scalar properties; extensions to vector-valued outputs, forces, excited-state observables, and kinetic parameters are left for future work.

Discussion & Conclusion

Augmenting geometry-based backbones with a compact set of string-level descriptors consistently improved predictions for electronically governed properties—such as dipole moment, frontier orbital energies, and the HOMO–LUMO gap—while leaving thermodynamic quantities largely unaffected or slightly degraded. Late-fusion architectures like SchNet and DimeNet++ integrated these auxiliary signals most effectively, indicating that global chemical context is best introduced after symmetry-aware message passing. In contrast, early fusion disrupted geometric invariances and led to diminished accuracy on shape-sensitive targets.

Feature selection patterns reinforce these observations. Although molecular weight was the most frequently selected descriptor, the highest average weights shifted toward chemically informative features—XLogP for LUMO energy and elemental formula for the HOMO–LUMO gap—highlighting the importance of polarity and composition in modulating electronic structure. For thermodynamic targets, no single textual feature exhibited dominant influence, emphasizing the need for richer, physics-grounded descriptors (e.g., force constants, vibrational modes) to achieve similar improvements.

In summary, this study demonstrates the potential of cooperative, agentic AI to selectively extract and integrate multimodal chemical knowledge. xChemAgents offers a step toward interpretable, physics-aware surrogates that can be extended beyond ground-state scalars to richer tasks in materials and molecular modeling.

Software and Data

The xChemAgents toolkit—including source code, curated textual descriptors, and the complete Selector–Validator implementation—is released under the MIT license and is available at the following link: https://github.com/KurbanIntelligenceLab/xChemAgents. All datasets and annotations are derived exclusively from open-access resources and are free of proprietary restrictions.

Impact Statement

xChemAgents integrates physics-aware language agents with GNNs to enable faster and more interpretable virtual screening, offering potential benefits for green catalyst design and drug discovery. By reducing descriptor overhead and producing transparent, logged rationales, the framework supports reproducibility and regulatory compliance. Key risks—including biased descriptor pools, overreliance on automated explanations, and the carbon footprint of training—can be mitigated through the use of balanced datasets, periodic rule audits, and energy-efficient fine-tuning strategies.

References

References

  1. [1]Kim, S., Chen, J., Cheng, T., Gindulyte, A., He, J., He, S., Li, Q., Shoemaker, B. A., Thiessen, P. A., Yu, B., et al. Pubchem 2025 update. Nucleic Acids Research, 53(D1):D1516–D1525, 2025.DOI
  2. [2]Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.arxiv.org/abs/1412.6980
  3. [3]Kohn, W. and Sham, L. J. Self-consistent equations including exchange and correlation effects. Physical review, 140(4A):A1133, 1965.DOI
  4. [5]Lin, L., Lu, J., and Ying, L. Numerical methods for kohn–sham density functional theory. Acta Numerica, 28:405–539, 2019.DOI
  5. [6]Luo, X., Wang, Z., Wang, Q., Lv, J., Wang, L., Wang, Y., and Ma, Y. Crystalflow: A flow-based generative model for crystalline materials. arXiv preprint arXiv:2412.11693, 2024.DOI
  6. [7]Michel, F., Ferber, J., and Drogoul, A. Multi-agent systems and simulation: A survey from the agent community’s perspective. In Multi-Agent Systems, pp. 17–66. CRC Press, 2018.
  7. [8]Orio, M., Pantazis, D. A., and Neese, F. Density functional theory. Photosynthesis Research, 102:443–453, 2009.DOI
  8. [9]Oviedo, F., Ferres, J. L., Buonassisi, T., and Butler, K. T. Interpretable and explainable machine learning for materials science and chemistry. Accounts of Materials Research, 3(6):597–607, 2022.
  9. [10]Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., and Lerer, A. Automatic differentiation in pytorch. In Advances in Neural Information Processing Systems Workshop on Autodiff, 2017.
  10. [11]Polat, C., Kurban, M., and Kurban, H. Multimodal neural network-based predictive modeling of nanoparticle properties from pure compounds. Machine Learning: Science and Technology, 5(4):045062, 2024.DOI
  11. [12]Polat, C., Kurban, H., and Kurban, M. Quantumshellnet: ground-state eigenvalue prediction of materials using electronic shell structures and fermionic properties via convolutions. Computational Materials Science, 246:113366, 2025a.DOI
  12. [13]Polat, C., Kurban, H., Serpedin, E., and Kurban, M. Understanding the capabilities of molecular graph neural networks in materials science through multimodal learning and physical context encoding, 2025b. URL https://arxiv.org/abs/2505.12137.
  13. [14]Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp. 8748–8763. PMLR, 2021.arxiv.org/abs/2103.00020
  14. [15]Ramakrishnan, R., Dral, P. O., Rupp, M., and von Lilienfeld, O. A. Quantum chemistry structures and properties of 134 kilo molecules. Scientific Data, 1, 2014.
  15. [16]Reiser, P., Neubert, M., Eberhard, A., Torresi, L., Zhou, C., Shao, C., Metni, H., van Hoesel, C., Schopmans, H., Sommer, T., et al. Graph neural networks for materials science and chemistry. Communications Materials, 3(1):93, 2022.arxiv.org/abs/2208.09481
  16. [17]Ruddigkeit, L., Van Deursen, R., Blum, L. C., and Reymond, J.-L. Enumeration of 166 billion organic small molecules in the chemical universe database gdb-17. Journal of chemical information and modeling, 52(11):2864–2875, 2012.
  17. [18]Satorras, V. G., Hoogeboom, E., and Welling, M. E (n) equivariant graph neural networks. In International conference on machine learning, pp. 9323–9332. PMLR, 2021.
  18. [19]Schütt, K., Unke, O., and Gastegger, M. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In International Conference on Machine Learning, pp. 9377–9388. PMLR, 2021.arxiv.org/abs/2102.03150
  19. [20]Song, T., Luo, M., Zhang, X., Chen, L., Huang, Y., Cao, J., Zhu, Q., Liu, D., Zhang, B., Zou, G., et al. A multiagent-driven robotic ai chemist enabling autonomous chemical research on demand. Journal of the American Chemical Society, 147(15):12534–12545, 2025.DOI
  20. [21]Thölke, P. and De Fabritiis, G. Torchmd-net: equivariant transformers for neural network based molecular potentials. arXiv preprint arXiv:2202.02541, 2022.arxiv.org/abs/2202.02541
  21. [22]Van der Hoek, W. and Wooldridge, M. Multi-agent systems. Foundations of Artificial Intelligence, 3:887–928, 2008.
  22. [23]Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., et al. Huggingface’s transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019.arxiv.org/abs/1910.03771
  23. [24]Xie, J. and Liu, C.-C. Multi-agent systems and their applications. Journal of International Council on Electrical Engineering, 7(1):188–197, 2017.DOI
  24. [25]Xie, T., Fu, X., Ganea, O.-E., Barzilay, R., and Jaakkola, T. Crystal diffusion variational autoencoder for periodic material generation. arXiv preprint arXiv:2110.06197, 2021.arxiv.org/abs/2110.06197
  25. [26]Zhang, J., Wang, Z., Zhu, H., Liu, J., Lin, Q., and Cambria, E. Mars: A multi-agent framework incorporating socratic guidance for automated prompt optimization. arXiv preprint arXiv:2503.16874, 2025.

Paper details

vlnc:W_rb576dqm
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents