Property prediction of Mn-Co-Ni NTC thermistors by direct learning from XRD patterns

Negative temperature coefficient (NTC) thermistors are indispensable for applications such as temperature sensing, monitoring, and overheating protection in electronic devices. For practical use, precise control of the B constant and electronic resistance (LogR) is required; however, these parameters are strongly affected by compositional changes, crystal phase transitions, and microstructural effects, making optimization difficult. In this study, we applied machine learning directly to X-ray diffraction (XRD) patterns of 50 Mn–Co–Ni-based NTC thermistors to predict B constant and LogR values. A one-dimensional convolutional neural network (Baseline model) achieved RMSE = 417 K and R2 = 0.54 for the B constant, and RMSE = 0.34 log kΩ mm and R2 = 0.85 for LogR. When compositional descriptors (591 dimensions) were combined with convolutional features (Baseline with descriptors model), accuracy further improved, yielding RMSE = 337 K and R2 = 0.70 for the B constant, and RMSE = 0.29 log kΩ mm and R2 = 0.89 for LogR. SHAP analysis provided input-level structural interpretability: Mn-related descriptors contributed positively to the B constant, diffraction peaks related to cubic spinel structures were associated with reduced resistance, whereas peaks from rock-salt NiO increased resistance. These results demonstrate that integrating direct learning of XRD patterns with compositional descriptors provides improved predictive accuracy and partial insight into composition – structure – property relationships in NTC thermistors. This study demonstrates that direct convolutional learning from experimental XRD patterns enables accurate and physically interpretable prediction of thermistor properties, even in small, multiphase datasets.