13 December 2023 in npj Computational Materials
In recent times, transformer networks have achieved state-of-the-art performance in a wide range of natural language processing tasks. Here we present a workflow based on the fine-tuning of BERT models for different downstream tasks, which results in the automated extraction of structured information from unstructured natural language in scientific literature. Contrary to existing methods for the automated extraction of structured compound-property relations from similar sources, our workflow does not rely on the definition of intricate grammar rules. Hence, it can be adapted to a new task without requiring extensive implementation efforts and knowledge. We test our data-extraction workflow by automatically generating a database for Curie temperatures and one for band gaps. These are then compared with manually curated datasets and with those obtained with a state-of-the-art rule-based method. Furthermore, in order to showcase the practical utility of the automatically extracted data in a material-design workflow, we employ them to construct machine-learning models to predict Curie temperatures and band gaps. In general, we find that, although more noisy, automatically extracted datasets can grow fast in volume and that such volume partially compensates for the inaccuracy in downstream tasks.
30 June 2022 in Physical Review B
We introduce a translational and rotational invariant local representation for vector fields, which can be employed in the construction of machine learning energy models of solids and molecules. This allows us to describe, on the same footing, the energy fluctuations due to the atomic motion, the longitudinal and transverse excitations of the vector field, and their mutual interplay. The formalism can then be applied to physical systems where the total energy is determined by a vector density, as in the case of magnetism. Our representation is constructed over the power spectrum of the combined angular momentum describing the local atomic positions and the vector field, and it can be used in conjunction with different machine learning schemes and data taken from accurate ab initio electronic structure theories. We demonstrate the descriptive power of our representation for a range of classical spin Hamiltonian and machine learning algorithms. In particular, we construct energy models based on both linear Ridge regression, as in conventional spectral neighbor analysis potentials, and the Gaussian approximation. These are both built to represent a Heisenberg-type Hamiltonian including a longitudinal energy term and spin-lattice coupling.
11 September 2023 in Physical Review B
We introduce a compact cluster expansion method constructed over Jacobi and Legendre polynomials to generate highly accurate and flexible machine-learning force fields. The constituent many-body contributions are separated, interpretable, and adaptable to replicate the physical knowledge of the system. In fact, the flexibility introduced by the use of the Jacobi polynomials allows us to impose, in a natural way, constraints and symmetries to the cluster expansion. This has the effect of reducing the number of parameters needed for the fit and of enforcing desired behaviors of the potential. For instance, we show that our Jacobi-Legendre cluster expansion can be designed to generate potentials with a repulsive tail at short interatomic distances, without the need of imposing any external function. Our method is here continuously compared with available machine-learning potential schemes, such as the atomic cluster expansion and potentials built over the bispectrum. As an example, we construct a Jacobi-Legendre potential for carbon by training a slim and accurate model capable of describing crystalline graphite and diamond, as well as liquid and amorphous elemental carbon.
20 April 2021 in Physical Review Materials
The and GeTe compounds are of interest for applications in phase change memories. In the reset process of the memory the crystal is rapidly brought above the melting temperature by Joule heating and then the liquid phase rapidly cools down leading to the formation of the amorphous phase. Since the liquid above is metallic and the amorphous phase is semiconducting a semiconductor-to-metal transition occurs in the supercooled liquid. Based on density functional simulations, we estimated the metal-semiconductor transition temperature by monitoring the opening of a band gap in the supercooled liquid phase. Due to previous evidence on the importance of the van der Waals (vdW) interaction in describing the liquid phase of these materials, we used both the revised Vydrov-van Voorhis functional which includes vdW nonlocal interactions and the Perdew-Burke-Ernzerhof functional without vdW corrections. The estimated is about 100–150 K higher with the former than with the latter framework for both compounds. By including vdW interactions the estimated is closer to than to the glass transition temperature for both systems. The analysis of the structural properties as a function of temperature suggests a correlation between the metal-semiconductor transition and a Peierls distortion. However, the data support more a continuous structural transformation than the presence of a first order liquid-liquid phase change associated the metal-semiconductor transition.
25 January 2024 in Journal of Chemical Information and Modeling
In the search for novel intermetallic ternary alloys, much of the effort goes into performing a large number of ab initio calculations covering a wide range of compositions and structures. These are essential to building a reliable convex hull diagram. While density functional theory (DFT) provides accurate predictions for many systems, its computational overheads set a throughput limit on the number of hypothetical phases that can be probed. Here, we demonstrate how an ensemble of machine-learning (ML) spectral neighbor-analysis potentials (SNAPs) can be integrated into a workflow for the construction of accurate ternary convex hull diagrams, highlighting regions that are fertile for materials discovery. Our workflow relies on using available binary-alloy data both to train the SNAP models and to create prototypes for ternary phases. From the prototype structures, all unique ternary decorations are created and used to form a pool of candidate compounds. The SNAPs ensemble is then used to prerelax the structures and screen the most favorable prototypes before using DFT to build the final phase diagram. As constructed, the proposed workflow relies on no extra first-principles data to train the ML surrogate model and yields a DFT-level accurate convex hull. We demonstrate its efficacy by investigating the Cu-Ag-Au and Mo-Ta-W ternary systems.
21 April 2020 in The Journal of Physical Chemistry C