Dissolved gas analysis (DGA) is widely used for condition assessment of oil-immersed power transformers, but public-data artificial intelligence (AI) studies can overstate evidential reliability when labels are heterogeneous, samples overlap across benchmarks, or random splits ignore asset and time structure. This study presents an evidence-hierarchy audit framework for trustworthy AI-assisted transformer DGA assessment. The framework separates public-label benchmark agreement, transparent weak-label reproduction, leakage-sensitive validation, and future field-confirmed validity; combines exact gas-label fingerprinting, tree-ensemble baselines, Shapley additive explanations (SHAP) attribution, and a Mechanism-Consistency Score (MCS); and uses confidence-MCS logic as an expert-review trigger. On the Kaggle five-gas benchmark, LightGBM, Random Forest, and XGBoost achieved mean Macro-F1 values of 0.8982, 0.8884, and 0.8878, while exact fingerprinting showed that all 2321 alan-456 samples overlapped with Kaggle-derived fingerprints. After overlap removal, cross-dataset Macro-F1 decreased to 0.4688–0.5015. On the Lewis weak-label data, transformer-grouped validation reduced Macro-F1 from 0.9992 to 0.8691, showing sensitivity to asset-level splitting assumptions. MCS analysis identified class-dependent mechanism consistency and low-MCS review signals under five-gas inputs. The results support public-benchmark reliability and evidence-quality auditing rather than field-confirmed diagnostic performance.

Material defects govern the performance limits, degradation pathways, and service lifetimes of functional and structural materials, yet their predictive modeling remains difficult because local electronic reconstruction, long-range interactions, rare kinetic events, and finite-temperature effects are coupled across scales. First-principles calculations provide reliable descriptions of defect cores and charge states but are limited by system size, time scale, and sampling cost, whereas empirical molecular dynamics (MD) can access larger systems but often lacks transferability in reconstructed or chemically complex defect environments. This review examines machine learning (ML) for defect modeling when model choice, training data, and validation criteria are matched to the physical problem. We discuss machine-learning interatomic potentials (MLIPs), sparse Gaussian-process and committee-machine models, active-learning workflows, high-throughput defect screening, and ML-assisted multiscale simulations that connect atomistic calculations with longer-timescale modeling. Particular attention is given to the distinct requirements of point defects, dislocations, surfaces, grain boundaries, and interfaces, including configurational sampling, long-range interactions, charge response, and uncertainty control. Representative applications in semiconductors, halide perovskites, solid electrolytes, battery cathodes, irradiation damage, and catalytic nanoparticles show where ML expands the feasible scientific questions and where first-principles benchmarks or experimental validation remain essential. Reliable ML-assisted defect modeling ultimately depends on validating the physical quantities that control the relevant defect mechanism, rather than relying on aggregate model accuracy alone.

To address the challenges of data sparsity and model generalization in small-sample material optimization for flame-retardant resin matrices in composites, this study proposes a “parameterization-visualization” dual-driven strategy. This approach establishes an efficient, standardized pathway for AI-driven material development without relying on black-box algorithms. First, bivariate experimental data involving 10-(2,5-dihydroxyphenyl)-10H-9-oxa-10-phospha-phenanthrene-10-oxide (DOPO-HQ) and a self-developed R-powder (composed of Y₂O₃-stabilized borophosphate glass, zinc borate, and SiO₂) were utilized to generate a high-resolution performance matrix via contour interpolation, thereby transforming sparse experimental points into continuous performance surfaces. Second, a customizable weighted total scoring function was introduced to map multi-objective optimization onto a visualized, single-objective comprehensive scoring field. This not only enabled the precise identification of compliant variable intervals and the global optimum—converting discrete data into a continuous parameter space—but also provided structured training datasets for future machine learning models. Guided by this strategy and leveraging the core AI model of Surrogate-Based Optimization (SBO), the optimal 48.3D/63.9R formulation was rapidly identified with only a limited number of trials. Driven by a synergistic dual-mode mechanism—gas-phase radical quenching by 9,10-dihydro-9-oxa-10-phosphaphenanthrene-10-oxide (DOPO) and condensed-phase ceramic glaze formation by molten R-powder—this system achieved superior processability and mechanical integrity while meeting the impregnation and curing demands of large-tow carbon fiber composites. Beyond demonstrating exceptional gas-phase/condensed-phase synergistic flame retardancy, this work validates the proposed parameterized optimization path as a critical bridge connecting small-sample experiments with future large-scale AI optimization. It offers a novel methodological paradigm for accelerating the intelligent design of high-performance flame-retardant resins tailored for large-tow composite materials, effectively overcoming current limitations in data scarcity and model adaptability.