Smarter Models for Stronger Decisions
in Science and Engineering

Technometrics is a journal of statistics for the physical, chemical, and engineering sciences. Published quarterly by the American Society for Quality and the American Statistical Association, it offers papers describing new statistical techniques and innovative applications of known statistical methods. It also publishes expository papers on statistical methods and papers dealing with the philosophy and problems of applying statistical methods.

The most recent issue of Technometrics features 19 research articles that represent a unified toolkit that can be synthesized into the following five major themes:

1. Gaussian Process Enhancements
Several studies address the computational limits of Gaussian processes in handling large-scale or nonstandard data. Researchers introduced random Fourier features to scale GPs for mixed qualitative-quantitative inputs. They also developed “double emulators” for simulators with nonstationary output space grounding and used singular value decomposition to create multi-output GPs for high-dimensional spatiotemporal flow fields. Additionally, the GEMSS subsampling methodology was proposed to improve GP predictive accuracy in unexplored regions while mitigating numerical instability.

2. Out-of-Distribution and Anomaly Detection
Detecting structural deviations is a critical focus. For safety-critical AI, one study leverages multi-class GPs to quantify logit uncertainty for out-of-distribution detection using only in-distribution data. In manufacturing, a novel generative adversarial approach uses limited real OOD samples to explore unseen defect spaces. For networked systems, the TopoDepth framework combines persistent homology and functional data depth to detect anomalies in weighted dynamic networks with varying nodes and edges.

3. Industrial Process Monitoring and Diagnostics
To handle complex streaming data, researchers developed INPOM, a control chart using a latent tensor GP with mixed effects to monitor spatial profiles without requiring complete data arrays. For high-dimensional processes under resource constraints, a monotonic neural network strategy adaptively selects critical data streams for observation, reducing detection delays. Finally, a hierarchical Bayesian framework integrating Cox proportional hazards and multi-output GPs was introduced to jointly predict multi-mode failures and remaining useful life.

4. Advanced Regression, Sampling, and Dimensionality Reduction
Methodological improvements include cellPCA, a robust principal component analysis method that simultaneously handles casewise, cellwise, and missing outliers. For interaction modeling, an adaptive cmenet method selects conditional main effects under generalized linear models. Additional innovations extend MCMC samplers to handle complex linear and nonlinear inequality constraints in multivariate normal distributions, while other innovations introduce parallel flats structures to simplify model selection in split-plot fractional factorial designs.

5. Multi-Fidelity and Applied Predictive Modeling
In applied settings, multi-fidelity parameter estimation methods were developed to leverage low-fidelity data alongside high-fidelity variables, focusing on extreme value analysis. Real-world applications include a discrete generalized Pareto and quantile regression framework to forecast weather-related electricity network faults and a Bayesian regression model optimized via simulated annealing to estimate energy consumption for electric buses.

From engineering and manufacturing to energy networks and computer simulations, the core objective of these methodologies is to make statistical learning more computationally efficient, robust to anomalies, and interpretable when traditional models fail.

Visit the ASA Member Hub for details about accessing Technometrics and other ASA journals.