From Experiments to Expertise: Scientific Knowledge Consolidation for AI-Driven Computational Research

In the field of computational materials science, AI applications have made significant progress. With the development of large language models (LLMs) in recent years, AI's ability to autonomously plan, execute, and interpret complex simulation tasks has greatly improved. However, existing AI systems lack knowledge accumulation and consolidation between executions, limiting their research capabilities. Traditional AI systems often treat each computation as an independent task, neglecting the valuable experience and knowledge accumulated between different computational tasks. This issue of knowledge isolation is particularly pronounced in computational materials science, where research often requires long-term experiential accumulation and knowledge consolidation.

Existing AI systems in computational materials science primarily focus on single execution tasks, lacking knowledge accumulation and consolidation. The isolation of knowledge between computation tasks prevents AI from effectively utilizing previous experience to optimize current tasks. This lack of knowledge consolidation not only limits AI's research capabilities but also leads to inefficiency and inaccuracy in complex simulation tasks. Additionally, AI systems often rely on pre-trained datasets and cannot apply accumulated knowledge to unfamiliar materials, further limiting their application in materials science research.

QMatSuite addresses the issue of knowledge isolation in AI-driven computational materials science by providing a persistent scientific memory system. Its core innovations include:

• �� Recording and retrieving computational results with full provenance: QMatSuite enables the sharing and application of scientific knowledge between different computational tasks through detailed recording and retrieval mechanisms.

• �� Dedicated reflection sessions to correct erroneous findings and synthesize observations: The platform provides dedicated reflection sessions, allowing AI to self-correct and optimize, paralleling human cognitive rhythms.

• �� Cross-compound pattern recognition: Through pattern recognition mechanisms, QMatSuite can identify and apply cross-compound knowledge patterns, enhancing AI's ability to apply accumulated knowledge to unfamiliar materials.

QMatSuite's core methodology includes the following steps:

• �� Recording computational results: At the end of each computational task, the platform automatically records the results and their provenance, ensuring the accumulation of scientific knowledge.

• �� Retrieving knowledge: Before starting a new computational task, the platform automatically retrieves previous computational results to help AI optimize the current task.

• �� Reflection sessions: In dedicated reflection sessions, the platform automatically identifies and corrects erroneous findings and synthesizes observations into cross-compound knowledge patterns.

• �� Pattern recognition: Through pattern recognition mechanisms, the platform can identify and apply cross-compound knowledge patterns, enhancing AI's ability to apply accumulated knowledge to unfamiliar materials.

In the experimental design, QMatSuite's knowledge consolidation capabilities were validated through a benchmark test on a six-step quantum-mechanical simulation workflow. The experiments involved 135 autonomous solid-state calculations and 98 molecular geometry optimizations, covering six material categories and 98 molecular geometry optimizations. The experiments used two different AI models and three simulation engines and were validated through a complex six-step anomalous Hall conductivity workflow. The results showed that QMatSuite effectively reduces reasoning overhead, improves computational accuracy, and applies accumulated knowledge to unfamiliar materials.

The experimental results showed that QMatSuite effectively reduces reasoning overhead and improves computational accuracy. In the benchmark test on a six-step quantum-mechanical simulation workflow, accumulated knowledge reduced reasoning overhead by 67% and improved accuracy from 47% to just 3% deviation from literature. When applied to an unfamiliar material, deviation was only 1% with zero pipeline failures. Additionally, QMatSuite demonstrated an 85.2% autonomous completion rate across 135 autonomous solid-state calculations and 98 molecular geometry optimizations. Lattice constants for 114 materials agree with experimental results at a mean absolute error (MAE) of 1.02%.

QMatSuite's application scenarios include:

• �� Application in computational materials science: Through knowledge consolidation, QMatSuite enhances AI efficiency and accuracy in complex simulation tasks, applicable to various materials science research.

• �� Application to unfamiliar materials: Through cross-compound knowledge pattern recognition, QMatSuite enables the application of accumulated knowledge to unfamiliar materials, enhancing AI's research capabilities.

• �� Application in other scientific fields: QMatSuite's persistent scientific memory system can be extended to other scientific fields requiring long-term experiential accumulation, aiding AI in achieving knowledge consolidation in these fields.

Despite significant progress in knowledge consolidation, QMatSuite still has some limitations. First, the platform may face timeouts when handling complex materials, particularly in topological insulators and layered transition metal compounds. Second, the platform's knowledge consolidation mechanism relies on the quality of reflection sessions, which, if insufficient, may lead to the accumulation of erroneous knowledge. Additionally, current knowledge consolidation primarily relies on existing computational results and has not yet achieved predictive capabilities for emerging materials. Future research directions include expanding QMatSuite's knowledge consolidation capabilities to handle more types of materials and simulation engines and enhancing the intelligence of reflection sessions.