Property Optimization of Mn–Ni–Al Oxide Thermistors via an Integrated Framework of Efficient Composition Exploration and Image Analysis

For negative‑temperature‑coefficient thermistors, resistance R and the B constant must be precisely controlled, yet they depend sensitively on composition, crystal phases, and microstructure. This study targets Mn–Ni–Al oxides (Al ≤ 30 mol%) and builds an informatics workflow integrating: (i) exhaustive synthesis guided by K‑means clustering, (ii) outlier‑oriented sampling by the BoundLess Objective‑free eXploration algorithm (BLOX), (iii) machine‑learning regression with composition‑histogram descriptors, and (iv) deep learning on SEM/EDS images. 25 cluster‑representative compositions are synthesized and evaluated for logarithmic R (log R) and B constant. BLOX, combining Random Forest with Stein novelty, then prioritizes 18 Mn/Ni‑edge compositions where properties may change abruptly, yielding a 43‑sample database that efficiently covers outlier regions. Regression attains root mean square error (RMSE) < 0.63 log (kΩ mm) and R 2 > 0.7 for log R, whereas B constant remains hard to predict owing to outliers. Feeding SEM/EDS images to a convolutional neural network lowers the RMSE of log R to 0.49, evidencing substantive contributions from microstructure and elemental distributions. Self‑attention visualization indicates that Mn spinel phases are critical electron‑hopping pathways. The workflow offers a paradigm for accurate prediction and optimization of thermistor properties and functional ceramics.