ARIA: Causal-Aware Materials Discovery

• ARIA is a causal-aware framework that integrates structured literature and mechanistic reasoning to improve LLM-based predictions in materials discovery.
• It routes queries through a three-tier cascade—direct causal reasoning, physics-informed analogy, and parametric fallback—to ensure robust evidence and mechanistic completeness.
• Empirical evaluations on 2D materials show ARIA improves prediction accuracy by up to 51%, enhancing interpretability and reliability in both forward and inverse design tasks.

ARIA (Autonomous Reasoning Intelligence for Accelerated Materials Discovery) is a causal-aware framework that integrates mechanistic knowledge from structured literature sources to guide LLMs in materials science applications. The architecture is designed to mitigate deficiencies in standard LLM-augmented reasoning—chiefly, "contextual tunneling" in which models over-anchor on specific, sometimes incomplete, pieces of retrieved evidence and fail to provide globally consistent, physically valid predictions. ARIA routes queries through a three-tiered cascade conditioned on the presence of mechanistically complete paths in a large knowledge graph (KG) assembled from the scientific literature. Its application to forward prediction and inverse design in the domain of two-dimensional (2D) materials demonstrates large improvements in physical validity, interpretability, and reliability over baselines (Cao et al., 21 Jun 2026).

1. Causal-Aware Reasoning and Contextual Tunneling

ARIA is motivated by the observation that LLMs augmented via retrieval simply patch together fragmented evidence, often missing critical mechanistic links such as full process–structure–property (PSP) chains. This can lead to the "contextual tunneling" problem: models myopically focus on partial or recent retrievals (e.g., a single Processing→Structure snippet), ignoring global constraints and causality, which results in physically implausible or incomplete outputs.

ARIA resolves this by enforcing a mechanistic completeness criterion over the PSP knowledge graph. Only when a full, high-confidence causal chain (P→S→P) matching the query is available does the system permit direct evidence-based reasoning. Otherwise, it seeks validated analogies or, if necessary, falls back on general physical and parametric priors with an explicit disclaimer.

2. Formal Mechanistic Completeness and Inference Cascade

For a query $q$ (consisting of synthesis method, material, doping, and property target), ARIA defines mechanistic completeness as follows. Let $G = (V, E)$ be the PSP knowledge graph derived from scientific literature, with $V$ partitioned into processing ($V_P$), structure ($V_S$), and property ($V_{Prop}$) nodes, and each edge $e \in E$ assigned a confidence $w(e) \in [0,1]$. Query $q$ is "PSP-complete" if there exists a three-hop path $v_P \rightarrow v_S \rightarrow v_{Prop}$, matching $G = (V, E)$0's entities, with all $G = (V, E)$1 (typically $G = (V, E)$2).

ARIA routes queries through an inference cascade:

1. Tier 1 (Direct Causal Reasoning): If a PSP-complete path exists, assemble all such chains and their provenance, and prompt the LLM to use only these mechanisms for prediction. Outputs include citations and confidence sourced from the KG.
2. Tier 2 (Physics-Informed Analogy): If no direct PSP-complete chain exists, retrieve top-$G = (V, E)$3 similar materials (by semantic, categorical, and numerical similarity threshold $G = (V, E)$4), adapt their mechanisms via element substitution and thermodynamic checks, and prompt the LLM for analogical justification with medium confidence flags.
3. Tier 3 (Parametric Fallback): If neither direct nor analogous evidence is available, inform the LLM that no mechanistic evidence exists and instruct it to proceed on general, low-confidence physical priors. Output is explicitly marked as speculative.

This structured gating constrains LLM outputs, ensuring that auditable causal traces and provenance are provided whenever possible.

3. Knowledge Graph Construction and Statistics

The ARIA knowledge graph is constructed via automated extraction from 1,531 full-text papers and 10,000 abstracts (2005–2023) on 2D materials. Extracted using Qwen2.5:7B, the KG comprises 2,839 manually filtered and confidence-weighted PSP relations, split as 2,271 for training and 568 held out for evaluation.

• Nodes:
  • $G = (V, E)$5 — Synthesis methods (e.g., chemical vapor deposition, hydrothermal, sol–gel)
  • $G = (V, E)$6 — Structures (phase, crystal symmetry, defect density)
  • $G = (V, E)$7 — Properties (band gap, conductivity, absorbance, thermal conductivity)
• Edges:
  • Directed, weighted, and supported by provenance (article ID, sentence), with two relations: P→S and S→Prop.
  • All extractions are filtered to ensure thermodynamic and stoichiometric validity; each edge keeps its extraction source for maximum auditability.

4. Quantitative Evaluation and Benchmarking

ARIA's efficacy is demonstrated on forward prediction (synthesis → structure/property) and inverse design (target property → synthesis protocol) over 149 expert-validated 2D materials cases (117 in-domain, 32 out-of-domain).

Key metrics include:

• Scientific Accuracy: Adherence to physical law
• Functional Equivalence: Does predicted synthesis achieve target property?
• Completeness: All process steps and parameters specified
• Interpretability: Narrative clarity and traceability to literature

All are aggregated by harmonic mean as the "overall score."

Performance summary (mean ± std):

Method	Task	In-Domain	Out-of-Domain
Baseline LLM	Forward	0.340±0.033	0.313±0.047
Naive KG+LLM	Forward	0.337±0.027	–
ARIA-CORE (static)	Forward	0.410±0.024	–
ARIA-FULL (+search)	Forward	0.512±0.039	0.513±0.040
Baseline LLM	Inverse	0.345±0.018	–
ARIA-CORE	Inverse	0.454±0.015	–
ARIA-FULL	Inverse	0.498±0.034	0.513±0.040

ARIA's improvements over baseline LLMs are $G = (V, E)$8 in overall score in both forward and inverse tasks; ARIA-FULL outperforms adaptive retrieval-augmented generation (Self-RAG) by $G = (V, E)$9 (forward, ID/OOD) and $V$0 (inverse, ID/OOD), robustly matching or exceeding generalization in held-out material categories (Cao et al., 21 Jun 2026).

5. Auditable Causal Traces and Exemplars

ARIA outputs are grounded in explicit, traceable PSP reasoning. Examples:

• Direct (Tier 1):
  • Query: Hydrothermal synthesis of Ni-doped MoS$V$1—expected conductivity?
  • Path 1: Hydrothermal → crystallinity↑ → conductivity↑ (w=0.82)
  • Path 2: Hydrothermal → defect density↓ → mobility↑ (w=0.78)
  • Final answer references both mechanisms, returns value with citations.
• Analogy (Tier 2):
  • Query: Sol–gel synthesis of V-doped WS$V$2, optical absorbance?
  • Analog candidate: Cr-doped WS$V$3 ($V$4), structural and valence similarity confirmed.
  • Causal path adapted, prediction justified by analogy and physical similarity.
• Fallback (Tier 3):
  • Query: Hydrothermal synthesis of 2D topological insulator XZ$V$5–Mo system.
  • No direct or analogical evidence. Output marks the speculative nature and bases estimate on general kinetic and thermodynamic reasoning.

This explicit reasoning and provenance differentiation enable interpretability, transparency, and trustworthiness essential for scientific discovery.

6. Significance and Implications

ARIA's causal-aware paradigm for LLM reasoning ensures that predictions in materials discovery are physically grounded, interpretable, and reliable. By making mechanistic completeness a gating criterion, ARIA resists over-anchoring on incomplete or misleading evidence—a known failure mode in standard LLM+retrieval architectures. Its combination of knowledge graph structuring, threshold-based evidence selection, and audit-trail outputs provide a framework applicable to other domains where scientific or engineering predictions must be both correct and explainable. The significant improvements in evaluation metrics for both in-domain and out-of-domain cases further indicate ARIA's generalization and robustness benefits for autonomous scientific reasoning.