OptoMatKG: A Traceable Knowledge Graph Framework with Semantic Retrieval for Optoelectronic Materials Science
Abstract
Extracting and correlating knowledge from heterogeneous scientific publications poses a fundamental data engineering challenge for data-driven materials discovery. Traditional keyword search lacks semantic depth, while generic large language models (LLMs) are prone to hallucinated predictions. To address these limitations, we present OptoMatKG, a literature-scale data engineering system whose core contributions are threefold: (1) a domain-specific knowledge graph construction pipeline based on a nine-category fullworkflow ontology; the resulting graph comprises 2.62 million entities and 3.38 million typed relations; (2) a cascaded multi-stage retrieval pipeline that synergistically integrates LLM-based query normalization, semantic vector indexing and Materials Project pre-filtering to enhance metric-aware candidate ranking; and (3) an evidence-constrained auditing module that forces every output to anchor to DOI-supported reasoning chains, thereby enforcing end-to-end hallucination prevention. Using optoelectronic inverse design as a validation testbed, experiments verify that our entity extractor reaches a micro-averaged F1 of 0.971. Comprehensive tests over 2000 real-world queries further demonstrate that our system surpasses keyword-based and singlestage baselines in Recall@K and metric coverage. The system accurately quantifies multi-objective performance trade-offs while ensuring fully traceable inference. Although validated on photodetectors, the proposed solution provides a generalizable, pluggable engineering framework readily extendable to other functional material discovery pipelines.
Related
- A traceable agentic framework for chemistry research query and analysis with ThermoML database
- Drug Molecular and Route Similarity-Aware Retrieval-Augmented LLM Generation with Knowledge Graph for Semantic Reasoning
- XinPilot: a chemist-designed open‑source desktop workspace for traceable high-throughput synthesis and purity analysis
- TDEaseNexus: A Traceable, Open-Source AI Agent Platform for Top-Down Proteomics Data Analysis
- A Shared-Task-State-Driven Multi-Agent Framework for Knowledge Graph Construction in Substation Operation and Maintenance
BibTeX
@misc{ni2026optomatkg,
title = {{OptoMatKG: A Traceable Knowledge Graph Framework with Semantic Retrieval for Optoelectronic Materials Science}},
author = {Ben Ni and Lei Zhang},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008315/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008315/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - OptoMatKG: A Traceable Knowledge Graph Framework with Semantic Retrieval for Optoelectronic Materials Science
AU - Ben Ni
AU - Lei Zhang
PY - 2026
DA - 2026/09/04
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008315/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008315/v1
AB - Extracting and correlating knowledge from heterogeneous scientific publications poses a fundamental data engineering challenge for data-driven materials discovery. Traditional keyword search lacks semantic depth, while generic large language models (LLMs) are prone to hallucinated predictions. To address these limitations, we present OptoMatKG, a literature-scale data engineering system whose core contributions are threefold: (1) a domain-specific knowledge graph construction pipeline based on a nine-category fullworkflow ontology; the resulting graph comprises 2.62 million entities and 3.38 million typed relations; (2) a cascaded multi-stage retrieval pipeline that synergistically integrates LLM-based query normalization, semantic vector indexing and Materials Project pre-filtering to enhance metric-aware candidate ranking; and (3) an evidence-constrained auditing module that forces every output to anchor to DOI-supported reasoning chains, thereby enforcing end-to-end hallucination prevention. Using optoelectronic inverse design as a validation testbed, experiments verify that our entity extractor reaches a micro-averaged F1 of 0.971. Comprehensive tests over 2000 real-world queries further demonstrate that our system surpasses keyword-based and singlestage baselines in Recall@K and metric coverage. The system accurately quantifies multi-objective performance trade-offs while ensuring fully traceable inference. Although validated on photodetectors, the proposed solution provides a generalizable, pluggable engineering framework readily extendable to other functional material discovery pipelines.
ER -