Validating neural networks for spectroscopic classification on a universal synthetic dataset

Abstract

To aid the development of machine learning models for automated spectroscopic data classification, we created a universal synthetic dataset for the validation of their performance. The dataset mimics the characteristic appearance of experimental measurements from techniques such as X-ray diffraction, nuclear magnetic resonance, and Raman spectroscopy among others. We applied eight neural network architectures to classify artificial spectra, evaluating their ability to handle common experimental artifacts. While all models achieved over 98% accuracy on the synthetic dataset, misclassifications occurred when spectra had overlapping peaks or intensities. We found that non-linear activation functions, specifically ReLU in the fully-connected layers, were crucial for distinguishing between these classes, while adding more sophisticated components, such as residual blocks or normalization layers, provided no performance benefit. Based on these findings, we summarize key design principles for neural networks in spectroscopic data classification and publicly share all scripts used in this study.

Crystal Structure Assignment for Unknown Compounds from X-ray Diffraction Patterns with Deep Learning

“…Furthermore, they are often restricted to the classification of existing compounds and can hardly be generalized to compounds outside their databases. Although some recent studies succeeded in symmetry classification in a large chemical space, it is hardly feasible to utilize these prior models for phase identification because the phases/structures of all of the inorganic compounds are much more numerous than their space groups. − To solve these issues, it is required that the model be extensible according to the need of the task. The ResNet-based structure-type prediction protocol introduced in this work is developed based on the above idea, using a carefully designed parameter, reliability value, to tackle the challenge.…”

Accelerating Materials Discovery: Automated Identification of Prospects from X‐Ray Diffraction Data in Fast Screening Experiments

“…However, we conducted a recent study that revealed the lack of sensitivity with respect to identifying minor peaks in patterns for established network structures. [20] The detection of multi-phase peaks is of great importance for this work, necessitating modifications to the network architecture to improve performance in minor peak identification. Although single-phase structure peaks regularly occur in the training data, multiphase peaks are inserted at random positions, making them outliers from the expected results.…”

“…Second, an appropriate network structure is required to handle the peculiarities of the diffraction patterns. For instance, we evaluated commonly used neural network structures for the analysis of XRD in a recent study and identified deficiencies in detecting minor peaks in the diffraction patterns, [20] which extend to the recognition of multi-phase samples in the material discovery data. Additionally, amorphous phases have not been considered in prior works, so modifications to the architecture of established networks are necessary to handle such components.…”

Synthetic pre-training for neural-network interatomic potentials

“…Synthetic data are also beginning to be used in the physical sciences: Aty et al [14] studied synthetic data as a pre-training task for determining experimental lipid phase behaviour from small-angle x-ray scattering patterns [14]; Anker et al [15] explored the use of synthetic data in interpreting inelastic neutron-scattering data [15]; Schuetzke et al [16] used synthetic data to mimic the characteristic appearance of experimental measurements for several spectroscopy methods [16].…”