Machine learning-based automatic sample-changing device for HPGe gamma spectrometry

High-purity germanium (HPGe) gamma spectrometry is widely used in environmental and geoscientific measurements, but manual sample replacement remains inefficient, particularly for low-background samples that require long counting times. Fixed counting-time strategies often lead to over-measurement of high-activity samples and insufficient precision for low-activity samples. To address these problems, this study developed an automatic sample-changing device and introduced machine learning to predict sample measurement requirements, with the aim of improving measurement efficiency, reducing labor dependence, and supporting intelligent determination of sample-changing time.

The automatic sample-changing device was developed by integrating a mechanical execution subsystem, a pneumatic subsystem, an electrical control subsystem, and a host-computer communication subsystem. The system was designed to complete sample gripping, transfer, loading, and unloading in coordination with spectral acquisition. On this hardware platform, an intelligent workflow comprising real-time spectral acquisition, feature extraction, model judgment, and automatic sample-changing was established. Spectral parameters, including net peak area, peak area statistical uncertainty, relative error, count rate, activity, and minimum detectable activity, were calculated in real time. For model development, nine soil samples with different ¹³⁷Cs specific activities, collected from the black soil region of Northeast China, the Loess Plateau, the purple soil region of Southwest China, and the red soil region of Southeast China, were measured at multiple counting times, yielding 242 training records. Five regression models, namely Extra Trees, Random Forest, Gradient Boosting, support vector regression (SVR), and a BP-MLP neural network, were constructed and compared. Model performance was evaluated using leave-one-group-out cross-validation grouped by sample, with the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and R² as the main indicators. In addition, 378 fixed-time measurements from independent soil samples served as a supplementary validation dataset for the selected model.

The developed system achieved automatic sample replacement and continuous coordination of the measurement process, providing a practical basis for unattended HPGe gamma spectrometry. Among the tested models, the ensemble tree models outperformed SVR and BP-MLP, with Gradient Boosting showing the best overall predictive performance. Under leave-one-group-out cross-validation, its average test RMSE, MAE, MAPE, and R² were 3.60%, 2.08%, 13.92%, and 0.73%, respectively. Parameter sensitivity analysis indicated that model performance improved substantially as the number of base learners increased and stabilized when the number reached approximately 300. The predictions agreed well with the observed relative errors in both the training and test samples, indicating that the model effectively characterized the relationship between spectral features and measurement uncertainty. In the supplementary validation based on fixed-time measurements, the Gradient Boosting model achieved an RMSE of 2.20, an MAE of 1.82, a MAPE of 21.72%, and an R² of 0.45. When a 5% relative error threshold was used to back-calculate the target measurement time, the model-derived median target time was 32.27 h, compared with a theoretical median of 20.90 h, suggesting a relatively conservative prediction tendency.

The proposed machine learning-based automatic sample-changing device enables intelligent, automated control of HPGe gamma spectrometric measurements. The Gradient Boosting model effectively captured the variation of relative error with measurement conditions and offers a practical means of determining sample-changing time. Although the fixed-time validation data cannot serve as direct point-by-point ground truth for the true optimal measurement time, the results demonstrate the usefulness of the model for supplementary error prediction and time-control assistance. This study provides a practical reference for improving HPGe measurement efficiency, reducing reliance on manual operation, and promoting the intelligent operation of gamma spectrometric systems.