Hierarchical Multi-Agent Reasoning Systems for Knowledge-Grounded Material Substitution

Tarjei Paule Hage; Subhadeep Pal; Wei Lu; Yu-Chuan Hsu; Gayla Lyon; Jiezhu Jin; Markus J. Buehler

cond-mat.mtrl-sciOct 9, 2026

Hierarchical Multi-Agent Reasoning Systems for Knowledge-Grounded Material Substitution

Authors: Tarjei Paule Hage, Subhadeep Pal, Wei Lu, Yu-Chuan Hsu, Gayla Lyon, Jiezhu Jin, Markus J. Buehler

Organizations: Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA, USA · Department of Civil and Environmental Engineering, Massachusetts Institute of Technology, Cambridge, MA, USA · Saint-Gobain Research North America, Northborough, MA, USA · Department of Civil and Environmental Engineering, Department of Mechanical Engineering, Schwarzman College of Computing, Massachusetts Institute of Technology, Cambridge, MA, USA

Abstract

AI-assisted scientific discovery increasingly seeks to move beyond generating plausible ideas toward reproducible, constraint-aware workflows that can support practical engineering decisions. We present MARS, a hierarchical Multi-Agent Reasoning System for knowledge-grounded material substitution under scientific and manufacturing constraints. We apply MARS to replace per- and polyfluoroalkyl substances (PFAS), a demanding materials-design problem in which substitutes must satisfy functional performance, environmental, experimental, and processing requirements. MARS decomposes substitution into a closed-loop workflow comprising requirement analysis, grounded candidate generation, and manufacturability assessment with process design. The system integrates retrieval-augmented generation, knowledge-graph reasoning, a database of lab-available materials, persistent memory of intermediate results, and rejection-aware iterative feedback. In a PFAS-substitution case study, MARS generates structured material requirements, identifies an evidence-grounded substitute candidate, and develops a lab-scale manufacturing route. In a multi-seed ablation evaluation, MARS achieves the highest overall performance across all three subsystems, with gains that persist when the backbone model is held fixed. These results suggest that hierarchical decomposition, knowledge grounding, and manufacturability-aware refinement improve material-substitution reasoning beyond model capability alone.

Figures & tables

Figure 1: Overall architecture of the hierarchical multi-agent framework. High-level schematic of the proposed closed-loop hierarchical multi-agent system for PFAS material substitution. Dark gray elements indicate multi-agent systems, while light gray elements indicate intermediate results. From a natural-language query

Q

, the system first extracts required properties

W

and hard constraints

H

, then identifies a grounded substitute candidate

S

with supporting evidence

E

, and finally assesses lab-scale manufacturability to produce either a process recipe

P

or blocking constraints

C

. All intermediate results are stored in persistent memory, enabling iterative refinement until a candidate passes the framework’s internal feasibility assessment.

Figure 2: Knowledge sources used across the hierarchical multi-agent framework. Overview of the knowledge sources used across the three subsystems of the hierarchical multi-agent framework. System 1 is grounded in a curated PFAS research corpus and a domain-specific KG, System 2 uses complementary KGs, patent data, and a database of lab-available materials to identify substitute candidates, and System 3 draws on process-relevant textual sources for manufacturability assessment and process design. Here, the database-stack symbols denote text vector databases, the graph symbols denote KGs, and the folder symbol denotes the custom-built internal materials database.

Figure 3: Schematic of the model-proposed substitute material. Schematic cross-section of the substitute material proposed by MARS for the THV case study, consisting of a polyolefin elastomer base, an EVOH gas-barrier layer, and an OTS oleophobic surface coating. Layer thicknesses are not drawn to scale. The schematic represents the model-proposed design; no physical sample was fabricated in the present study.

Figure 4: Closed-loop candidate refinement in the THV case study. Rejection-aware search trace for the THV substitution query, illustrating closed-loop interaction between System 2 (material feasibility screening) and System 3 (manufacturability assessment). Material substitution requirements produced by System 1 are passed to System 2, which iteratively generates and screens candidates, rejecting those that violate hard constraints, for example by containing PFAS. Candidates that pass System 2 screening are forwarded to System 3 for manufacturability evaluation. When System 3 rejects a candidate, structured feedback is returned to System 2, which initiates another round of candidate generation and screening. Solid arrows indicate the main workflow, while dashed arrows indicate internal rejections within System 2. The trace terminates when System 3 assigns a candidate the manufacturable verdict.

Configuration	MAS	RAG/KG	IR	Description
MARS	✓	✓	✓	Full framework using gpt-oss-20b , combining multi-agent decomposition, RAG/Graph-RAG grounding, and closed-loop iterative refinement.
MARS ( gpt-5.6-sol )	✓	✓	✓	Full MARS framework with gpt-5.6-sol as the backbone of all agents.
Ablation 3	✓	$\times$	$\times$	Three-stage gpt-oss-20b pipeline with subsystem-level decomposition but without RAG/Graph-RAG or iterative refinement.
Ablation 2	$\times$	✓	$\times$	Single-call gpt-oss-20b baseline with static RAG/Graph-RAG grounding but without iterative refinement.
Ablation 1	$\times$	$\times$	$\times$	Single-call gpt-oss-20b baseline without retrieval, graph grounding, or iterative refinement.
Ablation 1 ( gpt-5.6-sol )	$\times$	$\times$	$\times$	Single-call gpt-5.6-sol baseline without retrieval, graph grounding, or iterative refinement.

Table 1: Summary of the framework configurations evaluated in the ablation study. MAS denotes multi-agent system, RAG/KG denotes RAG/Graph-RAG integration, and IR denotes iterative refinement based on internal critique.

Figure 5: Evaluation performance across framework configurations. Performance comparison across the evaluated framework configurations.(a) Overall average score across the twelve evaluation criteria for all six configurations. (b) Subsystem-level scores for Systems 1–3, corresponding to requirement analysis, candidate generation, and manufacturability assessment, respectively. (c) Criterion-level comparison of the two MARS backbones across the three subsystems. Criteria are completeness (Compl), scientific correctness (Sci), evidence grounding (Grnd), investigation quality (Inv), requirement alignment (Algn), adversarial verification (Verif), candidate anchoring (Anchor), adaptive reasoning (Adapt), route coherence (Route), operational practicality (Prac), material-process compatibility (Compat), and feasibility-verdict grounding (Verd). Bars show mean judge scores across repeated runs, with error bars indicating standard deviations. For MARS ( gpt-5.6-sol ), the overall score in panel (a) and the System 3 score in panel (b) are calculated over four route-producing runs because one run terminated before System 3 was invoked. In panel (c), asterisks denote criteria for which one MARS run was unavailable or the corresponding capability was unexercised and was therefore excluded from the criterion mean. gpt-6-astra is used as the LLM judge.

Pipeline	System 1	System 2	System 3	Total
MARS ( gpt-5.6-sol )	$6.72\pm 0.77$	$6.26\pm 0.38$	$7.25\pm 0.37^{*}$	6.78
MARS ( gpt-oss-20b )	$5.85\pm 0.48$	$5.22\pm 0.35$	$5.35\pm 0.41$	5.47
1 LLM Call ( gpt-5.6-sol )	$4.00\pm 0.25$	$2.17\pm 0.11$	$4.47\pm 0.06$	3.55
1 LLM Call (w/ RAG/KG)	$4.53\pm 0.53$	$1.70\pm 0.19$	$4.10\pm 0.31$	3.44
1 LLM Call (No RAG/KG)	$3.40\pm 0.31$	$1.73\pm 0.06$	$3.85\pm 0.48$	2.99
3 LLM Calls (No RAG/KG)	$3.45\pm 0.23$	$2.05\pm 0.14$	$2.42\pm 1.28$	2.64

Table 2: Evaluation scores for each configuration on a 0–10 scale. Subsystem scores are reported as mean

\pm

standard deviation over five runs. Structurally absent capabilities score zero; unexercised MARS criteria are excluded. ∗ System 3 is averaged over four route-producing runs.

Figure 6: System 2 to System 3 feedback ablation. Isolation of System 3-to-System 2 feedback on the reported THV query. (a) Mean subsystem scores when System 3 blocking constraints are propagated to System 2 or withheld. System 1 is identical in both conditions. (b) Criterion-level comparison for Systems 2 and 3. Feedback produces only modest changes in System 2 scores but substantially improves all System 3 feasibility-related criteria. Bars show means over five judge repetitions; error bars indicate standard deviations.

RAG Source	Source Corpus Size	VectorDB File Size
PFAS research corpus	4,716 papers	1.3GB
Patent corpus	13,654 patents	69MB
Material specification sheets	20 documents	6.5MB
Polymer science textbooks	35 textbooks	260MB

Table 3: Summary of the text-based retrieval sources used in the study. These sources support RAG across the three subsystems of the hierarchical multi-agent framework.

Knowledge Graph	Source Corpus	Nodes	Edges	Graph File Size	Embedding File Size
PFAS	4,612 papers	144,335	459,248	137MB	432MB
Material Properties	62,999 abstracts	317,800	774,009	229MB	481MB
Patents	13,654 patents	199,008	578,361	154MB	594MB

Table 4: Summary of the three KGs used in the study. The PFAS-specific KG and the Material Properties KG are taken from [ 49 ] , while the Patents KG is newly constructed in this work.

Materials Science

Materials science studies how composition, structure, processing, and defects determine material properties and how to design useful materials.

See all 301 papers

ProvMind: Provenance-grounded reasoning for materials synthesis
May 27, 2026Yiming Zhang, Ryo Tamura, Koji TsudaReasoning Evaluation Materials Science
Graduate School of Frontier Sciences, The University of Tokyo, Chiba, Japan. · Center for Basic Research on Materials, National Institute for Materials Science, Tsukuba, Ibaraki, Japan. · RIKEN Center for Advanced Intelligence Project, Tokyo, Japan.
ARIA: A Causal-Aware Framework for Rescuing LLM Reasoning in Trustworthy Materials Discovery
Jun 21, 2026Yi Cao, Liaoyaqi Wang, Jieneng Chen +3Knowledge Augmentation for Language Models Materials Property Prediction
Chemical and Biomolecular Engineering, Johns Hopkins University, Baltimore, MD, USA · Computer Science, Johns Hopkins University, Baltimore, MD, USA
Graph-Native Reinforcement Learning Enables Traceable Scientific Hypothesis Generation through Conceptual Recombination
Jul 1, 2026Subhadeep Pal, Shashwat Sourav, Tirthankar Ghosal +1LLM Reasoning with Graphs RL for Language Model Reasoning
Department of Civil and Environmental Engineering MASSACHUSETTS INSTITUTE OF TECHNOLOGY Cambridge, MA, USA · Department of Physics Washington University in St. Louis St. Louis, MO, USA · Computing and Computational Sciences Directorate Oak Ridge National Laboratory Oak Ridge, TN, USA +3

May 27, 2026J/K move · Enter open · S save

ProvMind: Provenance-grounded reasoning for materials synthesis

Yiming Zhang, Ryo Tamura, Koji Tsuda

Graduate School of Frontier Sciences, The University of Tokyo, Chiba, Japan. · Center for Basic Research on Materials, National Institute for Materials Science, Tsukuba, Ibaraki, Japan. · RIKEN Center for Advanced Intelligence Project, Tokyo, Japan.

Reasoning Evaluation Materials Science Structured Reasoning LLM Reasoning

Open paper AskPDF arXiv

Materials process optimization requires reasoning over routes, conditions, tools and causal dependencies, yet most computational formulations flatten synthesis procedures into text or ordered steps. We introduce MatProcBench, a provenance-grounded benchmark constructed from literature-mined MatPROV graphs, to evaluate seven process-reasoning tasks spanning route continuity, step-level variable inference and global causal consistency under both same-split and shift-aware evaluation, including a strict dual-OOD split that combines temporal and material-class shift. We further introduce ProvMind, a process-memory reasoning framework that retrieves analogous training processes, converts them into provenance-aware option-level compatibility scores, and uses a language model for constrained final decision making. ProvMind achieves 52.84% accuracy on the dual-OOD split, outperforming prompting, retrieval-augmented and supervised fine-tuning baselines.

Abstract

Materials Science

Explore similar work

ProvMind: Provenance-grounded reasoning for materials synthesis

ARIA: A Causal-Aware Framework for Rescuing LLM Reasoning in Trustworthy Materials Discovery

Graph-Native Reinforcement Learning Enables Traceable Scientific Hypothesis Generation through Conceptual Recombination