materials characterization

This work addresses the inaccuracy in hotspot prediction during transient thermal simulation of heterogeneous back-end-of-line (BEOL) structures in 3D chip stacks. To overcome this challenge, the authors propose a transient multiscale thermal analysis framework that extends multiscale homogenization—previously limited to steady-state scenarios—to transient thermal modeling. The method automatically extracts layout structures from GDSII/OASIS files to construct representative volume elements (RVEs), and under the assumption of temperature-independent material properties, it derives homogenized thermal conductivity and volumetric heat capacity, yielding an analytical expression for effective transient thermal conductivity. Validated on a 1 mm × 1 mm SoC model with 5 μm and 10 μm RVEs, the approach accurately captures transient thermal behavior at a time step of dt = 0.001, significantly enhancing the fidelity of thermal modeling for three-dimensional heterogeneous BEOL architectures.

Halide perovskite thermal stability is highly composition-dependent, yet its degradation mechanisms remain complex and poorly predictable. Method: We employ high-throughput in situ environmental photoluminescence (PL) characterization coupled with machine learning—specifically, a dual-mode XGBoost framework integrating composition-agnostic and composition-aware modeling. The methodology incorporates t-SNE/UMAP dimensionality reduction, feature importance analysis, and correlation heatmap visualization. Contribution/Results: We discover, for the first time, a significant negative correlation between Cs content and thermal stability—a counterintuitive trend. Our standardized, transferable ML pipeline achieves 85% PL stability prediction accuracy in the composition-agnostic mode and 75% in the composition-aware mode; Cs-related features dominate with 99% importance. This work establishes a generalizable analytical paradigm applicable to arbitrary perovskite systems and substantially accelerates the high-throughput screening of photovoltaic-grade stable perovskites.

Material discovery is hindered by the scarcity of large-scale, machine-readable experimental datasets linking crystal structures to thermoelectric properties. To address limitations of existing databases—including small size, heavy reliance on manual curation, and theoretical bias—this work introduces an LLM-based agent workflow that automatically extracts thermoelectric performance and crystal structure data from nearly 10,000 scientific publications. Our approach innovatively integrates dynamic token allocation, zero-shot multi-agent coordination, and conditional table parsing, achieving high extraction accuracy (F1 > 0.92) while reducing inference cost by 40%. We construct the largest LLM-curated thermoelectric dataset to date—comprising 27,822 temperature-resolved, unit-standardized records—and release an open-source platform supporting semantic search and interactive data export. This significantly enhances scalability and practicality for data-driven materials discovery.

This study addresses the challenge of performing in situ structural characterization of large-format combinatorial material libraries under high-temperature and controlled-atmosphere conditions, which existing high-throughput techniques struggle to achieve. The authors developed a wide-bore furnace capable of accommodating full 100 mm silicon wafers, integrated with synchrotron-based X-ray diffraction (XRD) and X-ray fluorescence (XRF), enabling in situ high-throughput analysis up to 735 °C across atmospheres ranging from nitrogen to pure oxygen. Combined with ternary oxide libraries fabricated via pulsed laser deposition and a custom MATLAB-based thermal expansion analysis program, this platform enabled, for the first time, wafer-scale in situ XRD characterization under realistic processing conditions. The approach overcomes limitations of conventional sample stages and reveals the inadequacy of Vegard’s law in predicting lattice behavior within high-entropy oxide systems.

Addressing urgent demands for novel functional materials in energy, microelectronics, and biomedical applications, this work tackles the challenge of material design optimization under high-dimensional, discrete, sparse, and fragmented data conditions. Method: We propose a tightly integrated hybrid modeling paradigm that synergistically combines first-principles calculations, multi-fidelity modeling, graph neural networks, Bayesian optimization, and knowledge-guided transfer learning to jointly predict material properties and infer viable synthesis pathways. Contribution/Results: We systematically identify critical bottlenecks—including material data quality deficiencies and performance–application mismatches—previously uncharacterized. Evaluated on MIT’s benchmark material design task, our framework achieves efficient convergence from tens of thousands of candidates to a high-confidence top-10 set. It successfully discovers multiple new metal–insulator transition materials and proposes experimentally verifiable synthesis routes.

This study addresses the challenge of integrating and interpreting heterogeneous, multiscale characterization data of carbon nanotube (CNT) thin films arising from microstructural variations. To this end, the authors propose an interpretable multimodal machine learning framework—the first application of multimodal explainable AI to multiscale materials characterization. The approach fuses multimodal data, including morphological features extracted from SEM images via binarization, skeletonization, and network analysis, along with Raman spectroscopy, specific surface area, and surface resistivity measurements. By combining UMAP for dimensionality reduction and XGBoost for nonlinear regression, the framework enables end-to-end performance prediction. Under leave-one-out cross-validation, the model achieves optimal predictive accuracy. Feature importance analysis reveals that surface resistivity is primarily governed by inter-junction transport length, crystallinity, and network connectivity, while specific surface area is dominated by junction density and pore size, thereby elucidating the microstructure–property relationships in CNT films.

This work proposes a robust machine learning workflow to address the limited generalization capability that often undermines reliable prediction of the thermoelectric figure of merit (zT) in Half-Heusler compounds. By employing PCA-based unbiased data splitting to ensure chemical representativeness of the test set, the framework integrates k-best feature selection, Bayesian hyperparameter optimization, and multi-model ensembling. Physical interpretability is enhanced through SHAP and SISSO analyses, shifting the evaluation focus from conventional metrics toward generalization reliability. The approach identifies A-site doping concentration and enthalpy of vaporization as key descriptors for zT and efficiently screens over 660 million candidate structures, yielding multiple stable, high-zT materials that validate the framework’s effectiveness and practical utility.

This study addresses the challenge of accurately extracting microstructural parameters from scanning electron microscopy (SEM) images of sintered ceramics, where conventional segmentation methods struggle due to unimodal gray-level histograms and overlapping intensity distributions between grains and pores. To overcome this limitation, the authors propose an automated analysis pipeline that integrates topological filtering with the Sauvola local thresholding method. The approach first employs topological filtering to effectively suppress noise and then applies the Sauvola algorithm to achieve high-precision segmentation, enabling automatic quantification of porosity, solid-phase fraction, and grain/pore size distributions. Experimental results on samples sintered at 1200 °C and 1400 °C demonstrate segmentation Intersection-over-Union (IoU) scores of 95.14% and 99.85%, respectively, significantly outperforming manual analysis and achieving notable advances in both accuracy and efficiency.

The discovery of high-performance thermoelectric materials (with high ZT) remains inefficient due to the lack of quantitative design principles. Method: This work proposes κₗ/κ ≈ 0.5—the ratio of lattice to total thermal conductivity—as the first quantitative instantiation of the “phonon-glass electron-crystal” concept. A dual-task machine learning model jointly predicts κ and κₗ/κ, trained on 71,000 high-quality experimental and computational data points. It integrates physics-informed feature engineering and high-throughput screening. Contribution/Results: Applied to >100,000 inorganic compounds, the model identifies 2,522 candidates with ultralow κ. Experimental validation via doping and alloying confirms that the κₗ/κ ≈ 0.5 criterion effectively guides thermoelectric performance optimization. This work establishes a paradigm shift in thermoelectric materials design—from empirical heuristics to data-driven, quantitative prediction.

This study addresses the challenge of defect visibility fluctuation in time/frequency domains during infrared thermography post-processing, which hinders automated unsupervised analysis due to conventional evaluation metrics’ reliance on prior knowledge of defect locations or defect-free reference regions. To overcome this limitation, the work proposes a data-driven, spatial-prior-free approach that adapts the concept of representative elementary volume into two dimensions for infrared thermography and integrates three no-reference metrics: a homogeneity index (HI) based on local intensity distribution deviation, a representative elementary area (REA) derived from Minkowski functionals, and a geometric-topological total variation energy (TVE). Validated on pulsed thermography sequences of a carbon-fiber-reinforced polymer plate containing six artificial defects, the method enables robust and unbiased frame ranking, offering a reliable basis for automated defect detection through optimal image selection.