
Machine-learned interatomic potential methods have revolutionized the field of materials science. These methods use artificial neural networks to predict the behavior of materials at the atomic level.
By leveraging large datasets and complex algorithms, researchers can create highly accurate models of material properties. This enables the design of new materials with tailored properties.
Machine-learned interatomic potentials have been applied to a wide range of materials, including metals, semiconductors, and polymers. This has led to breakthroughs in fields such as energy storage, electronics, and biomedical devices.
The accuracy of machine-learned interatomic potentials has been demonstrated to be comparable to, or even surpass, traditional empirical potential methods.
Related reading: VACUUM
Machine Learning
Machine-learned interatomic potentials, or MLIPs, can be trained to have a number of a priori unknown parameters that are found by requiring that key quantities are close to their quantum mechanics (QM) counterparts on the training set.
Active learning is used to construct the training set, allowing the MLIP to be trained on the fly. This is particularly useful in long molecular dynamics simulations, where the MLIP can calculate forces and occasionally request QM data for fitting.
The reliability of passive learning on the fly is limited, but active learning can provide perfectly reliable results.
Training of MLIPs
Training of MLIPs is a crucial step in machine learning. It involves finding a number of a priori unknown parameters that are required to make the key quantities, such as total energy and forces on atoms, close to their QM counterparts.
These parameters are found by requiring that the key quantities are close to their QM counterparts on the training set. The training set can be constructed by active learning that allows training a MLIP on the fly.
Active learning is particularly useful in the context of long molecular dynamics, where MLIP is employed for calculating the forces and from time to time requests QM data for fitting. This approach is of high practical importance.
A comparison of different learning approaches shows that active learning on the fly (LOTF) is perfectly reliable, whereas passive learning on the fly has limited reliability. This highlights the importance of active learning in training MLIPs.
For another approach, see: Active Learning (machine Learning)
Message-Passing Networks
Message-Passing Networks are a type of neural network architecture that allows for more efficient computation and better scalability.
They work by breaking down complex neural networks into smaller, more manageable pieces, called layers, which communicate with each other through a process called message passing.
This approach can lead to significant improvements in training speed and model performance.
The key insight behind message-passing networks is that many neural network tasks can be decomposed into smaller, more local computations that can be easily parallelized.
For example, consider a neural network that is trying to classify images. The network can be broken down into smaller layers that process different aspects of the image, such as edges, textures, and shapes.
By passing messages between these layers, the network can learn to represent the image in a more efficient and effective way.
Message-passing networks have been shown to be particularly effective in tasks such as graph neural networks and molecular property prediction.
These tasks involve processing complex, high-dimensional data that can be difficult to work with using traditional neural network architectures.
By using message-passing networks, researchers have been able to achieve state-of-the-art results in these areas.
Take a look at this: Machine Learning in Computer Networks
Applications and Methods
Machine-learned interatomic potential (MLIP) is a powerful tool with a wide range of applications.
MLIPs are successfully applied in research projects and applications of atomistic modeling, with a rapidly growing number of such projects.
Crystal structure prediction is one of the many applications of MLIP, where it helps identify the most stable crystal structures of materials.
The training set for MLIP includes a variety of crystalline and liquid configurations, with 24 crystalline and 76 fully or partly liquid configurations.
By leveraging these configurations, MLIP can accurately estimate atomization energies of molecules and point defects diffusion.
Emerging Methods
Gaussian Approximation Potential (GAP) is a machine-learned interatomic potential that combines compact descriptors of local atomic environments with Gaussian process regression to machine learn the potential energy surface of a given system.
This approach has been successfully applied to various systems, including elemental systems like Carbon, Silicon, Phosphorus, and Tungsten, as well as multicomponent systems like Ge2Sb2Te5 and austenitic stainless steel.

The GAP framework has been used to develop a number of Machine-Learned Interatomic Potentials (MLIPs) for different materials, demonstrating its effectiveness in modeling complex systems.
Machine learning is a key aspect of GAP, allowing it to learn the potential energy surface of a system from a set of training data.
Moment Tensor Potentials (MTPs) are another family of local MLIPs that use a linear regression model with a specific set of basis functions to describe neighborhoods of atoms.
These functions are invariant with respect to rotations, reflection, translation, and permutation of atoms of the same chemical type, making them a powerful tool for modeling complex systems.
MTPs have been shown to have a high performance compared to Density Functional Theory (DFT), and their computational complexity scales linearly with the number of atoms.
The performance of MTPs can be further improved by systematically increasing the number of parameters, allowing them to approximate the behavior of any Quantum Mechanics (QM) model.
Here are some key differences between GAP and MTPs:
- GAP uses Gaussian process regression, while MTPs use linear regression.
- GAP is more general-purpose, while MTPs are designed for specific types of systems.
- GAP has a higher computational complexity, while MTPs scale linearly with the number of atoms.
Symbolic Regression
Symbolic regression is a type of machine learning that involves finding a mathematical equation that best fits a given set of data. This method is particularly useful for modeling complex systems and identifying underlying patterns.
One of the key benefits of symbolic regression is its ability to provide interpretable results, making it easier to understand the relationships between variables. This is especially important in fields like engineering and economics where decision-makers need to be able to explain their findings to stakeholders.
Symbolic regression can be used to model a wide range of phenomena, from the behavior of physical systems to the performance of financial portfolios. In one example, researchers used symbolic regression to model the relationship between temperature and the viscosity of a fluid, resulting in a highly accurate equation that could be used for predictive purposes.
The process of symbolic regression typically involves using an evolutionary algorithm to search for the best equation that fits the data. This algorithm can be tuned to favor certain types of equations or to prioritize simplicity over accuracy.
Symbolic regression can also be used to identify the most important variables in a dataset, which can be a major advantage in situations where data is limited or noisy. By identifying the key drivers of a system, researchers and practitioners can focus their efforts on the most critical areas.
Applications of Mlip

MLIPs are successfully applied in a variety of research projects and applications, including crystal structure prediction and finding stable phases of alloys.
The number of such projects is rapidly growing, with automated alloy phase diagrams being one of the notable applications.
Estimating atomization energies of molecules is another area where MLIPs have shown promise.
Point defects diffusion is also being studied using MLIPs, which is an important aspect of materials science research.
An illustration of configurations selected into the training set by active learning shows the energy range of crystalline configurations after relaxation, with 24 crystalline configurations and 76 fully or partly liquid configurations featured in the training set.
Units
In most machine-learned inter-atomic potential studies, the unit system is eV for energy, Angstrom for distances, and mass in atomic units. This has become the default unit system for these models.
These algorithms don't care about the units, but they can only be trained on one set of units, so it's essential to know what units they're using. The unit system is not limited by the models themselves.
Times are often measured on the femtosecond scale. This is also the standard unit system used within the Atomic Simulation Environment (ASE).
We'll use the pint library to manage our units throughout the tutorial.
Featured Images: pexels.com





