2026/01/01 by Zelong Qiao, Run Jiang, Dapeng Cao
Computer Science · Materials Science · #Applications of artificial intelligence #Big Data and Digital Economy #Causal reasoning #Cognitive robotics #Human intelligence #Inference #Key (lock) #Machine Learning in Materials Science #Nanoporous metals and alloys #Path (computing) #Reinforcement learning #Verifiable secret sharing
paper · doi:10.1360/csb-2025-5797
openalex publication_date 2026/01/01 · crossref issued 2026/01/22 · crossref published 2026/01/22 · crossref published-online 2026/01/22 · crossref created 2026/05/08 · openalex created_date 2026/05/09 · openalex updated_date 2026/07/31 · crossref published-print 2026/08/01 · crossref deposited 2026/08/05 · crossref indexed 2026/08/05
<sec><p indent="0mm">Artificial intelligence (AI) is revolutionizing materials design by transitioning from mere data-driven tools to systems capable of logical reasoning. This paper discusses the evolutionary path of AI, highlighting a research framework transition from “direct data matching” to “indirect logical deduction” as a key milestone. We propose that this core criterion for categorizing AI levels should be “whether it possesses explainable causal reasoning and logical deduction capabilities.” It can be divided into two levels: broad AI, which excels in pattern recognition and statistical prediction, and Special AI, which demonstrates foundational cognitive abilities through interpretable inference. This distinction helps clarify the role of AI in scientific research, beyond functional tools, toward collaborative partners. </sec><sec> Based on technological advancements, we define Broad AI as the systems that primarily rely on direct data matching, such as deep learning models (e.g., AlphaFold and GNoME). These models act as complex “parameterized knowledge databases,” achieving high accuracy in specific tasks but limited to interpolation and extrapolation within training data distributions. In contrast, Special AI (e.g., OpenAI o1 and DeepSeek-R1) embodies indirect logical deduction, enabling causal inference and adversarial thinking through process-supervised reinforcement learning. This means a fundamental transition from “know what” to “know why”, with Special AI capable of generating verifiable reasoning chains and adapting to open-ended problems. </sec><sec> In catalytic materials design, we propose an Agent framework built on descriptors, which can efficiently reveal the structure-property relationship and key variables linking micro-scale structure and macro-scale performance. Descriptors provide interpretable, computable features that integrate domain knowledge with data-driven optimization. The Agent operates via a “perception-decision-execution-learning” closed-loop workflow, enhancing research efficiency and interpretability. For instance, the perception module gathers R&D requirements, while decision-making leverages descriptor databases for prior knowledge injection, avoiding blind searches. </sec><sec> We further speculate that an “Agent clusters + Master controller” architecture could represent the mature form of general AI. Here, the Master controller employs a dynamic weight allocation (inspired by the concept of hybrid functionals in density functional theory) to coordinate specialized Agents (e.g., for demand parsing or theoretical computation). Weights are adjusted based on real-time feedback, task phases, and historical data, enabling adaptive decision-making. This system would drive AI from an auxiliary tool into a “robotic scientist”, although these challenges, like multi-source knowledge fusion and computational resources, still exist. </sec><sec> In short, this systematic exploration underscores the transformation of AI in materials design. By emphasizing the deduction capability as a criterion, Broad AI (direct data matching) and Special AI (indirect logic deduction) are distinguished. The descriptor-based Agent framework promises cross-disciplinary innovation. As key technologies mature, AI is poised to transition from an auxiliary tool into a “robotic scientist”, driving research paradigm innovation in catalytic materials design. </sec>