Agentic schema-guided extraction of materials process knowledge from scientific literature

Sameer Sadruddin; Jennifer D'Souza

cs.CLOct 5, 2026

Agentic schema-guided extraction of materials process knowledge from scientific literature

Authors: Sameer Sadruddin, Jennifer D'Souza

Organizations: TIB Leibniz Information Centre for Science and Technology, Hannover, Germany.

Abstract

Materials literature contains detailed experimental knowledge, but procedures, chemical entities and measurements remain difficult to aggregate because they are reported in heterogeneous forms and depend on process-specific context. We present SciKGExtract, a schema-guided framework that combines large-language-model extraction with chemical normalization and agent-based evaluation and refinement before knowledge-graph integration. We evaluate the framework on 176 atomic-layer-deposition papers describing zinc oxide (ZnO) and indium--gallium--zinc oxide (IGZO), together with an expert-annotated full-schema subset. PubChem normalization improves exact-match extraction F1 for every tested model. For ZnO, the best F1 increases from 0.591 for direct normalized extraction to 0.805 with agentic refinement, whereas the best IGZO result is 0.344, revealing the greater difficulty of multicomponent supercycle processes. Evaluation against a deeply nested schema containing 65 experimental properties and 155 quantitative measurement nodes further exposes errors in process segmentation and numerical assignment. These results show that chemical canonicalization and targeted agentic verification provide complementary controls for converting complex materials literature into reusable, machine-actionable experimental knowledge.

Figures & tables

Figure 1: Partial structure of the ALD experimental process schema. The schema organizes an ALD process into nested branches for the deposition system, reactants, process parameters, material properties, device properties, and other experimental aspects. Quantitative observations are represented using QUDT-grounded quantity structures. The complete extraction target comprises 65 leaf properties and 155 QUDT-grounded quantity nodes. The complete schema is publicly available as an ORKG template at https://orkg.org/templates/R1434174 .

Figure 2: SciKGExtract orchestration workflow. The orchestrator coordinates Extract & Normalize, Reflect, Feedback and KG Integration agents. Structured process records can be evaluated by a single judge, multiple judges or evaluator–critic pairs and revised before accepted records are prepared for knowledge-graph integration.

Material	Model	Strategy	Unnorm. F1	Precision	Recall	Norm. F1	Accuracy
ZnO	GPT-4o	MJ2	0.591	0.658	0.741	0.697	0.83
ZnO	Llama 3.3 70B	Debate	0.761	0.780	0.831	0.805	0.91
ZnO	Mistral Large	MJ2	0.629	0.696	0.729	0.712	0.89
ZnO	Gemma 4 26B	MJ2	0.582	0.705	0.734	0.719	0.90
ZnO	Ministral 3B	MJ2	0.473	0.637	0.713	0.673	0.83
IGZO	GPT-4o	MJ2	0.202	0.239	0.304	0.268	0.82

Table 1: Best exact-match ALD result for each extraction model after PubChem normalization. Unnormalized F1 is shown for the same refinement configuration to isolate the effect of normalization. MJ2 and MJ3 denote multi-judge refinement with two and three judges, respectively.

Figure 3: PubChem-based normalization in SciKGExtract. Extracted chemical values are resolved through API and local lookup components, with an optional language-model disambiguation step for unresolved forms. The original extracted value is preserved and linked to resolved PubChem identifiers rather than replaced.

Material	Strategy	Precision	Recall	F1	Accuracy
ZnO	Direct	0.577	0.607	0.591	0.88
ZnO	Self-refinement	0.599	0.636	0.617	0.89
ZnO	Multi-judge, two judges	0.702	0.748	0.724	0.87
ZnO	Debate	0.780	0.831	0.805	0.91
IGZO	Direct	0.244	0.319	0.277	0.82
IGZO	Self-refinement	0.261	0.348	0.298	0.83

Table 2: Effect of refinement strategy for Llama 3.3 70B after chemical normalization. Debate-self uses GPT-5 as evaluator and critic; debate-cross uses GPT-5 as evaluator and DeepSeek R1 as critic.

		ZnO			IGZO
Model	Setting	P	R	F1	P	R	F1
GPT-4o	Before	0.42	0.39	0.41	0.50	0.49	0.48
GPT-4o	After	0.45	0.42	0.44	0.48	0.45	0.45
Mistral Large	Before	0.53	0.50	0.51	0.24	0.18	0.20
Mistral Large	After	0.54	0.51	0.53	0.37	0.30	0.32
Qwen3 30B	Before	0.42	0.39	0.40	0.43	0.42	0.41

Table 3: Exact-match full-schema extraction on expert-annotated ALD processes before and after one reflection–feedback cycle. This experiment uses the complete 65-leaf-property target and a separate controlled model sweep from the 176-paper refinement comparison.

Dataset	Model	Strategy	Precision	Recall	Exact F1
BioRED	Claude Sonnet 4.6	Debate-cross	0.3145	0.2580	0.2834
BC5CDR	Claude Sonnet 4.6	Multi-judge, two judges	0.6985	0.7542	0.7253

Table 4: Best exact-match transfer result observed on each biomedical benchmark.

Explore similar work

Mined from Scientific Literature: Process Schemas for Atomic Layer Deposition and Etching in Materials Science
Sep 14, 2026Sameer Sadruddin, Eleni Poupaki, Alex Watkins +5Materials Science Lithography
TIB Leibniz Information Centre for Science and Technology, Hannover 30167, Germany · Department of Applied Physics and Science Education, Eindhoven University of Technology, Eindhoven 5600 MB, The Netherlands · Department of Chemistry, University of Warwick, Coventry CV4 7AL, United Kingdom +1
SciSchema.org: A Multidisciplinary Collection of Schemas for Structured Scientific Process Descriptions
Jul 30, 2026Jennifer D'Souza, Sameer Sadruddin, Anisa Rula +23Schema Scientific Discovery
TIB Leibniz Information Centre for Science and Technology, Hannover, Germany · University of Brescia, Brescia, Italy · University of Alicante, Alicante, Spain +15
Agents-K1: Towards Agent-native Knowledge Orchestration
Jun 11, 2026Zongsheng Cao, Bihao Zhan, Jinxin Shi +23Knowledge Graphs Scientific Discovery
Shanghai Artificial Intelligence Laboratory · East China Normal University · Fudan University

Sep 14, 2026, cs.AIJ/K move · Enter open · S save

Mined from Scientific Literature: Process Schemas for Atomic Layer Deposition and Etching in Materials Science

Sameer Sadruddin, Eleni Poupaki, Alex Watkins, Bora Karasulu+4

TIB Leibniz Information Centre for Science and Technology, Hannover 30167, Germany · Department of Applied Physics and Science Education, Eindhoven University of Technology, Eindhoven 5600 MB, The Netherlands · Department of Chemistry, University of Warwick, Coventry CV4 7AL, United Kingdom+1

Materials Science Lithography Atoms Schema Materials Keyed Javascript Object Notation Record Yields+1

Open paper AskPDF arXiv

Atomic layer deposition (ALD) and atomic layer etching (ALE) are reported heterogeneously across experimental and simulation literature in materials science, hindering comparison and machine-actionable reuse. We present four domain-expert-reviewed JSON Schemas for ALD and ALE experimental and simulation processes. Curated with schema-miner and grounded in QUDT using schema-miner pro, the schemas structure materials, process conditions, configurations , and measured or predicted results. We compare their scope, structure, and semantic grounding, and demonstrate their use for schema-guided literature extraction and publication of structured records through ORKG templates.

Abstract

Explore similar work

Mined from Scientific Literature: Process Schemas for Atomic Layer Deposition and Etching in Materials Science

SciSchema.org: A Multidisciplinary Collection of Schemas for Structured Scientific Process Descriptions

Agents-K1: Towards Agent-native Knowledge Orchestration