Machine learning strategies for optimising frictional properties of materials

The programme aims at combining physics of friction - and materials science in general - with machine learning, in order to explore possibilities to efficiently optimise and predict materials properties, e.g. for applications where low friction is desired.
Aalto University/Pilvi Takala

The programme consists of a series of thematic workshops, visitors, and AScI fellows working together on physics and machine learning. It is a genuinely interdisciplinary programme exploiting the key strengths of Aalto University, e.g. materials research and computational science.

Principal investigators

Contact: [email protected]

Machine Learning for Friction Process

Controlling and reducing friction and wear between two surfaces in contact is of enormous practical importance, given that a significant fraction of the gross national product of the developed countries is wasted on friction and wear. A related and timely issue in nanotechnology is that due to continuing device miniaturization, friction imposes serious constraints on the performance and lifetime of various mechanical microdevices. At the same time, computational materials research is currently moving towards system scale thinking, where simulations of individual materials are no longer enough to meet the demands of device design. For computational studies of friction, the two solids in contact and the possible lubricant layer in between need to be considered in detail. However, simulating such complex systems at high accuracy is still extremely expensive, and the number of possible systems (i.e. different material/lubricant combinations) to consider is enormous. Therefore, novel computational strategies for “smart” optimization of the frictional properties of materials/lubricants are needed.

Machine learning strategies have been recently developed e.g. in order to accelerate materials property predictions. The key idea there is that discovering new materials with desired properties can be significantly expedited if one is able to effectively learn from the available knowledge and data. To this end, various machine or statistical learning methods, trained on data from numerical simulations, have been developed. The aim of the present project is to develop such strategies for optimization in general as an initial target. Some current work within Aalto University has already made progress in applying machine learning for efficient design of microelectronics interfaces and visualisation of complex systems. However, the approaches used and techniques developed have been very specific to each problem, and this project will generalise the methods for application to a much wider set of issues. This will touch on Aalto key strategic directions in energy research such as optimization of materials in plasmonics, catalysis, battery technologies and photovoltaics - and particularly the study of friction and wear.

In order to provide a coherent framework for the AScI theme, we will focus on investigation into frictional properties of materials and lubricants, with a strong link to leading experimentalist groups to test the resulting predictions. The key difference here as compared to the above-mentioned materials property predictions is that we are considering a specific phenomenon, i.e. friction, and plan to apply machine learning techniques (such as supervised and unsupervised learning) for optimization of frictional properties. We expect such a programme to have a significant potential for breakthroughs within tribology, to contribute to the development of novel strategies to reduce friction and wear in applications (mechanical microdevices, etc.), and to serve as a kickoff for a novel focused activity with the aim of networking across disciplines and internationally.

The project builds on some of the key strengths of Aalto, i.e. computational science and materials research. It has a distinct interdisciplinary character in that both physicists and computer scientists will be involved. It will further strengthen collaboration between two groups within the COMP CoE, i.e. the Surfaces and Interfaces at the Nanoscale (SIN) and Complex Systems and Materials (CSM). Moreover, the project is likely to gain synergies from an ongoing joint project of two CoEs at Aalto (COMP and COIN) on Bayesian construction of atomistic interaction potentials, and through our participation in the COST network “Understanding and Controlling Nano and Mesoscale Friction”.

Additional Research: Machine Learning for Atmosphere Science

Atmospheric aerosol particles are minute molecular agglomerations suspended in the air. These small particles take part in a multitude of physical processes some of which have very consequential societal impact. Aerosol particles affect the daily lives of millions of people in the form of deteriorated air quality: air pollution by particulate matter is one of the leading global causes of death and disability. Perhaps even more globally, atmospheric aerosol particles affect the global climate via aerosol-cloud interactions: aerosol particles often offer the needed pre-existing surface for cloud formation.

Atmospheric scientists are thus interested to understand how various processes modify the properties of the aerosol particles and especially how, why and when these particles form. However, due to the vast physical and chemical complexity of the atmosphere, the paths leading to atmospheric aerosol particle formation are not yet completely understood. One straightforward approach to overcome this issue is to gather more data.

Towards this end, several Stations for Measuring the forest Ecosystem-Atmosphere Relationships (SMEAR) were established in Finland. These stations measure numerous variables, including gases, meteorology, radiation, aerosol particles and soil fluxes. At SMEAR II in Hyytiälä forest alone, there are more than 200 observables in total since 1996 (i.e. millions in sample size), producing a big data.                                     

Machine Learning methods have been applied extensively by researchers and practitioners in order to discover patterns or predict outcomes from prior data, traditionally in the fields of image recognition, medical diagnosis, robotics, etc. Curiously, Machine Learning has not yet gained significant attention from the atmospheric science community. In this project, we will exploit the power of Machine Learning to mine the SMEAR database seeking to provide valuable information into atmospheric science. In order to provide a coherent framework for the AScI theme Machine Learning for Applied Physics, we will focus on developing Machine Learning strategies to deal with two particular research problems in this field.

First problem is related to determination of event and non-event days. Atmospheric scientists are interested in creating a database to study the cause and effects of atmosphere particle formation in the boreal forest. However, the current classification method is based on knowledge-based method, where scientists classify particle formation event/non-event days based on visualization scheme. The visualization-based classification may be subjective and may not be efficient for long period classification. Therefore, we will develop a Machine Learning model which can distinguish automatically event/non- event days probabilistically, so the users will not only be able to determine event/non-event days, but also to quantify the quality of classification outcomes.                 

In the second problem, we will investigate the key variables behind the atmospheric aerosol particle formation. This study is very intricate due to the complex physical and chemical processes involved. Therefore we will develop a “more complex” Machine Learning classifier involving measured data from SMEAR, to find relationships between these data sets and event classification dataset. This approach should be able to map the dependence of the various variables in relation to the formation event days and reveal the possible roles of previously overlooked physical parameters.

Additional Research: Machine Learning for Quantum Many-body Systems

Following Dirac's suggestion, we are currently applying machine-learning methods to the solution of quantum many-body problems. Since this is a complex problem, we start from the Hartree-Fock theory to keep it simple, and assume the wave-function of the n electrons to be a Slater determinant of single-particle molecular orbitals (MOs):

(kaavio)

and that the molecular orbitals are linear combinations of the B basis set atomic orbitals:
 

(kaavio)

The problem then reduces to calculating the Fock operator matrix elements, and diagonalise it to obtain eigenvalues (MO energies) and eigenvectors (linear combination coefficients). While the part of Hamiltonian (and Fock operator the as well) describing the kinetic energy and the electron-nuclei interactions can be calculated from the description of the system (atomic positions) and basis set, the electron-electron cannot. In order to estimate the electronic repulsion, we need to know where the electrons are. But to compute where they are we need to know their repulsion energy. The standard procedure is to start with a guessed wave-function and iteratively refine it until convergence. Each iteration requires O(N4) operations to build the operator, followed by O(N3) for diagonalisation. It is fascinating that the wave-function is entirely determined by the atomic positions and the total number of electrons, however it cannot be calculated directly from those.

Our goal is to machine-learn away this complexity. Currently, we are formulating a model to treat the arbitrary-size descriptor of the system, process it into a fixed size fingerprint from which matrix elements can be calculated. While the iterative approach is still required to converge the fingerprint, this is done entirely by a recurrent neural network, avoiding the O(N4) routine and the diagonalisation at each iteration.

Should our efforts be successful, it will become possible to calculate the wave-function, and thus all electronic properties, of larger systems with reduced computational effort. Moreover, this work can open up the possibility of training more advanced models with high level accurate quantum chemistry methods (MP2, coupled cluster and CI), thus providing us with an invaluable tool for academic and applied.

  • Published:
  • Updated:
Share
URL copied!