An auditable decision framework for donor–acceptor materials screening

Technical Summary: An Auditable Decision Framework for Donor–Acceptor Materials Screening

Problem Statement
Machine learning (ML) models can rapidly rank molecular candidates for donor–acceptor (D–A) materials, which are critical for organic electronics, photovoltaics, and photocatalysis. However, ranked predictions alone are insufficient to justify advancing candidates to costly experimental validation. A useful D–A pair must satisfy a coupled set of requirements: accessible excitation energies, long-lived charge-transfer states, suitable redox alignment, and specific aggregation behaviors. Current screening workflows often rely on single-property ranking or black-box predictions that fail to distinguish between high-performing candidates that are well-supported by training data and those that are uncertain or lie outside the model's learned chemical regimes. There is a need for a framework that treats the "molecular pair plus its evidence" as the screening object, enabling traceable, evidence-guided prioritization rather than simple property optimization.

Methodology
The authors propose an evidence-conditioned workflow that converts D–A pair predictions into auditable screening decisions. The system operates through three integrated layers:

1. Paired Surrogate Model:

   • Input Representation: Each candidate is treated as an ordered molecular pair (Side A: Donor, Side B: Acceptor) based on frontier orbital energy ordering. The input consists of 80 role-specific scalar descriptors (40 per side) and two monomer graphs (15 atom-level and 9 bond-level features per monomer).
   • Architecture: A two-stream cross-attention encoder processes the scalar descriptors, while an edge-aware graph branch encodes molecular topology. A gated multimodal fusion mechanism combines these into a shared latent vector.
   • Outputs: The model predicts eight continuous optical/electronic quantities, two multiclass excited-state character labels, and three binary aggregation-related outcomes.
   • Training: The model was trained on 2,440 ordered pairs using a monomer-disjoint split to prevent data leakage. It employs sharpness-aware minimization and local donor-acceptor augmentation.

2. Evidence Generation and Interpretation Layer:

   • Uncertainty Quantification: Monte Carlo dropout (16 passes) provides uncertainty estimates for regression (standard deviation), classification (predictive entropy), and binary outputs.
   • Local Explainability (XAI): For each query, the system generates a structured packet including:
     • Descriptor Attribution: Integrated gradients to identify influential scalar features.
     • Graph Saliency: Gradient-times-input analysis for atoms and bonds.
     • Perturbation Tests: Measuring output changes when removing specific modalities (descriptors vs. graphs) or masking specific molecular sides.
     • Latent-Neighbour Support: Assessing whether the query lies near learned chemical regimes using nearest-neighbor analysis in the latent space.
   • Global Audits: A reusable background context aggregates dataset-level attributions, side-responsibility trends, accumulated local effect (ALE) curves, and latent-regime prototypes/criticisms to provide a reference for interpreting local explanations.
   • Literature Retrieval: A hybrid search (BM25 + sentence-transformer) retrieves relevant passages from a curated corpus of 15 documents to provide chemical context.

3. Constrained Reporting Layer:

   • A chemistry-adapted Large Language Model (LLM), fine-tuned via 4-bit low-rank adaptation (LoRA) on OLMo-3-1025-7B, converts the fixed evidence packet into a chemist-facing report.
   • Constraint: The LLM is strictly prohibited from altering predicted values, confidence labels, or attribution outputs. Its role is solely to organize the evidence, state caveats, and recommend validation steps based on the provided data.

Key Contributions

• Auditable Decision Framework: The paper introduces a workflow that unifies prediction, uncertainty, local/global explanations, distributional support, and literature context into a single "evidence packet" for each D–A pair.
• Evidence-Conditioned Prioritization: Unlike traditional screening that ranks by a single metric, this framework uses a Pareto-based screen to balance predicted performance against model reliability (confidence, interpretability, and support).
• Role-Aware Multimodal Prediction: The architecture explicitly preserves molecular roles (Donor/Acceptor) and combines scalar chemistry with graph topology, allowing for the prediction of aggregate properties from isolated monomer information.
• Traceable Reporting: The use of a constrained LLM ensures that the final recommendation is a readable summary of the underlying evidence without introducing generative hallucinations or modifying the numerical data.

Results

• Predictive Performance: On the audited monomer-disjoint holdout set (combining validation and test partitions, totaling 1,069 pairs), the model achieved mean absolute errors of 0.07–0.30 for continuous outputs and accuracy ranging from 0.59 to 0.77 for categorical outputs.
• Local XAI Insights: Explanations revealed that different endpoints rely on different evidence streams. For instance, optical and electronic outputs were often weighted toward Side A in descriptors but Side B in graph signals, while state-character predictions were highly sensitive to graph information removal.
• Batch Screening: Applied to the held-out test split, a prespecified multi-objective rule identified 19 non-dominated candidates. The top-ranked candidate (d48_a13) balanced a predicted S1 oscillator strength of 0.0389 with high confidence (0.613) and no low-confidence targets.
• Decision Logic: The framework successfully distinguished between candidates that were "darker but less certain" (requiring targeted optical validation with uncertainty checks) and those that were "higher-brightness but higher-confidence" (suitable as lower-risk controls).

Significance and Claims
The authors claim that this work recasts materials screening from single-property ranking to evidence-guided prioritization. The primary significance lies in providing a traceable route to select D–A pairs for higher-level calculations and experimental testing. By treating the screening object as the "molecular pair together with the evidence needed to judge its predicted performance," the framework enables chemists to inspect, challenge, and act on recommendations.

The paper explicitly states that the workflow does not replace high-fidelity calculations or experiments but serves as a validation triage tool. It shifts the output from a single predicted score to an auditable candidate record that exposes which predictions are supported, which endpoints require caution, and which follow-up tests are most appropriate. The authors note that while the framework improves consistency and reduces ad hoc human selection bias, it remains bounded by the chemical space and descriptor quality of the training data, and mechanistic interpretations still require independent validation.