Synthesizing or studying certain materials in a laboratory setting often poses challenges on account of safety concerns, impractical experimental conditions, or cost constraints. In response, scientists are increasingly turning to deep learning methods which involve developing and training machine learning models to acknowledge patterns and relationships in data that include details about material properties, compositions, and behaviors. Using deep learning, scientists can quickly make predictions about material properties based on the fabric’s composition, structure, and other relevant features, discover potential candidates for further investigation, and optimize synthesis conditions.
Now, in a study published on 1 February 2024 within the International Union of Crystallography Journal (IUCrJ), Professor Takashiro Akitsu, Assistant Professor Daisuke Nakane, and Mr. Yuji Takiguchi from Tokyo University of Science (TUS) have used deep learning to predict single-molecule magnets (SMMs) from a pool of 20,000 metal complexes. This progressive strategy streamlines the fabric discovery process by minimizing the necessity for lengthy experiments.
Single-molecule magnets (SMMs) are metal complexes that exhibit magnetic rest behavior at the person molecule level, where magnetic moments undergo changes or rest over time. These materials have potential applications in the event of high-density memory, quantum molecular spintronic devices, and quantum computing devices. SMMs are characterised by having a high effective energy barrier (Ueff) for the magnetic moment to flip. Nevertheless, these values are typically within the range of tens to a whole bunch of Kelvins, making SMMs difficult to synthesize.
The researchers used deep-learning to discover the connection between molecular structures and SMM behavior in metal complexes with salen-type ligands. These metal complexes were chosen as they could be easily synthesized by complexing aldehydes and amines with various 3d and 4f metals. For the dataset, the researchers worked extensively to screen 800 papers from 2011 to 2021, collecting information on the crystal structure and determining if these complexes exhibited SMM behavior. Moreover, they obtained 3D structural details of the molecules from the Cambridge Structural Database.
The molecular structure of the complexes was represented using voxels or 3D pixels, where each element was assigned a singular RGB value. Subsequently, these voxel representations served as input to a 3D Convolutional Neural Network model based on the ResNet architecture. This model was specifically designed to categorise molecules as either SMMs or non-SMMs by analyzing their 3D molecular images.
When the model was trained on a dataset of crystal structures of metal complexes containing salen type complexes, it achieved a 70% accuracy rate in distinguishing between the 2 categories. When the model was tested on 20,000 crystal structures of metal complexes containing Schiff bases, it successfully discovered the metal complexes reported as single-molecule magnets. “That is the primary report of deep learning on the molecular structures of SMMs,” says Prof. Akitsu.
A lot of the anticipated SMM structures involved multinuclear dysprosium complexes, known for his or her high Ueff values. While this method simplifies the SMM discovery process, it will be important to notice that the model’s predictions are solely based on training data and don’t explicitly link chemical structures with their quantum chemical calculations, a preferred method in AI-assisted molecular design. Further experimental research is required to acquire the information of SMM behavior under uniform conditions.
Nevertheless, this simplified approach has its benefits. It reduces the necessity for complex computational calculations and avoids the difficult task of simulating magnetism. Prof. Akitsu concludes: “Adopting such an approach can guide the design of progressive molecules, bringing about significant savings in time, resources, and costs in the event of functional materials.”
