Deep learning sorts crystal structures for single-molecule magnet behaviour

The work

Technical · from the record

The authors assembled a dataset of salen-type metal complexes whose single-molecule magnet (SMM) behaviour had been reported in roughly 800 papers from 2011–2021, took the crystal structures from the Cambridge Structural Database, and encoded each molecule as a 64 × 64 × 64 voxel image coloured by element. A ResNet-style 3D convolutional neural network was trained to classify molecules as SMM or non-SMM and reached approximately 70% accuracy and AUC on the test split. Grad-CAM maps indicated that the model's contributions concentrated near the central metal. The trained model was then run over approximately 20 000 Schiff base crystal structures from the CSD, and the 10 most and 10 least SMM-like predictions were compared with the magnetic measurements in their original publications; several of the molecules predicted to be non-SMMs had never been measured by AC susceptibility.

How AI was used

A binary classifier was trained to decide, from 3D molecular structure alone, whether a salen-type metal complex shows single-molecule magnet behaviour. Labels came from the literature, with non-SMM defined as no magnetic relaxation in AC magnetic susceptibility measurements and zero-field results preferred where both zero-field and DC-field results existed. CIF files from the CSD were checked in Mercury and converted to XYZ files with crystal water and other non-pertinent molecules removed, then rendered as 64 × 64 × 64 × 3 voxel grids, 12 Å per side, with per-element RGB colour and sphere radius scaled to ionic radius. The 2:1 SMM:non-SMM imbalance was corrected by undersampling, and the data were split 6:2:2 into training, validation and test sets. The network was a 3D-CNN built from bottleneck ResNet blocks with pre-activation ReLU, a global average pooling layer into fully connected layers, batch normalisation, 20% dropout and a sigmoid output; it was trained for 1000 epochs with Adam/AMSGrad (ε = 1 × 10⁻⁷, β1 = 0.9, β2 = 0.999), a cosine learning-rate decay from 1 × 10⁻² to 1 × 10⁻⁵, batch size 8 and cross-entropy loss, with rotation, scaling and translation of the atomic coordinates as data augmentation. Grad-CAM was applied to the trained network to visualise which voxel regions drove predictions. The trained model was then applied, without further fitting, to approximately 20 000 Schiff base metal complex structures retrieved with ConQuest and processed with the CSD Python API, and the resulting scores were ranked to pick the 10 highest and 10 lowest for literature follow-up. Implementation used Python 3, TensorFlow 2.0, NumPy, SciPy and Open Babel.