Technology

We provide a theoretical regret analysis for Gaussian process Thompson sampling (GP-TS). By constructing a lower bound on the regret that grows polynomially with a certain probability, we show that logarithmic dependence in high-probability performance guarantees is generally unattainable. We further derive an upper bound based on the second moment of cumulative regret and an upper bound on expected regret with an introduced tolerance parameter. In addition, we establish a tight upper bound on cumulative regret under relaxed conditions for Matérn kernels, thereby strengthening the theoretical guarantees of GP-TS from multiple perspectives.

Elemental analysis of biofuels, including carbon, hydrogen, oxygen, and nitrogen (CHON), is essential for evaluating combustion quality and environmental impact. However, conventional destructive and time-consuming methods hinder rapid screening. In this study, we developed a non-destructive and rapid prediction framework that combines AggMap, which converts FTIR spectra into two-dimensional structures, with a hybrid convolutional neural network. The framework achieved high predictive accuracy for both elemental composition and higher heating value (HHV), enabling immediate fuel evaluation at the laboratory scale.

Conventional peak detection requires expert-driven manual parameter tuning for each type of spectrum, posing a major challenge for high-throughput analysis. In this study, we developed a method that automatically predicts optimal detection parameters from spectral features and achieved high accuracy in validation across diverse spectra. We also clarified that the effective parameter space, in which favourable detection results are obtained, varies depending on spectral characteristics. By ensuring that predicted parameters fall within this space, the method enables consistently accurate automated peak detection even for complex spectral shapes.

Conventional gas sensors face challenges in extracting detailed molecular information comparable to gas chromatography–mass spectrometry (GC-MS), resulting in limited interpretability and black-box characteristics. To address this issue, we developed a method that correlates principal component analysis (PCA)-compressed GC-MS data with QCM sensor time-series signals using a one-dimensional convolutional neural network. The method successfully reconstructed chemically interpretable two-dimensional mass spectrometry maps from sensor signals with high accuracy.

We demonstrated the effectiveness of a robust approach that improves the quality of experimental data and enables optimisation of the target yield by introducing a proprietary coefficient-based masking procedure.

Focusing on the importance of materials data infrastructure, we proposed a Materials Informatics process that integrates data quality assessment, feature selection, and optimisation. Through data quality control using proprietary metrics, data selection based on regression accuracy, and optimisation methods informed by feature contributions, the study demonstrates the potential to improve materials discovery performance and standardise the discovery process.

This review article examines individual authentication technologies based on odour components, namely volatile organic compounds (VOCs), contained in human skin emissions and breath. In addition to discussing the potential for identification based on chemical characteristics, the paper systematically summarises measurement methods and challenges for practical implementation, incorporating the authors’ research on sensor arrays and humidity mitigation technologies.

Real-world optimisation often requires black-box optimisation under the possibility of observation failures. To address such settings, this study proposes Failure-aware Gaussian Process UCB (F-GP-UCB). The method is established under the assumption that the optimum lies in the interior of the feasible region, and it is theoretically shown that the number of successful observations increases linearly. Based on this result, the study derives the first regret bound for this problem setting and confirms the effectiveness of the method across multiple benchmarks, including simulations that mimic materials synthesis experiments.

In real-world optimisation, it is often important to obtain the best possible outcome under randomness caused by environmental factors. This study formulates a risk-seeking optimisation problem for obtaining the best result within a fixed budget. Considering two settings, one in which environmental variables are observable and one in which they are not, the study proposes a new method, Kernel Explore-Then-Commit (Kernel-ETC). Theoretical regret bounds are provided, and the effectiveness of the method is demonstrated through experiments including hyperparameter tuning and polymer synthesis data.

This study addresses Bayesian optimisation problems in which observations of the objective function fail stochastically. Existing methods for this problem lack theoretical guarantees and may also be unstable in practice. We propose two methods: the first no-regret algorithm with theoretical guarantees for this setting, and a second method with strong practical performance. We also show that the latter can be modified to obtain theoretical guarantees. Experiments, including simulations that mimic quasicrystal synthesis, validate the effectiveness of both methods.