fields/materials-chemistry/model families
Multilayer perceptrons in materials and computational chemistry
How fully connected feedforward neural networks, the plainest kind of learned model, are used in materials science and computational chemistry to stand in for costly quantum-mechanical calculations.
How it is used
Materials science and computational chemistry try to work out how atoms arrange themselves and how those arrangements set the properties of a substance: whether it melts, how it conducts, how it deforms, what light it absorbs. The usual tool is quantum mechanics, solved approximately on a computer. These first-principles methods are accurate but slow, so they can only follow a small number of atoms for a short span of time. A multilayer perceptron is a simple neural network: layers of numbers multiplied by adjustable weights, with no built-in assumption about the structure of the problem. Here it is usually trained on quantum calculations and then asked to reproduce their answers much faster.
The commonest use is a learned description of the forces between atoms. The network is shown energies and forces computed from first principles for many atomic arrangements, and learns to predict them for new ones, which lets simulated atoms be moved about for far longer than the quantum method would allow. Networks are also used more narrowly, as a fitted stand-in for calculations that are expensive in a particular regime, or as a way of fitting the coefficients of an existing physical equation, with the equation's own terms written into the network's layers so that what is learned stays interpretable.
Why researchers use it
The appeal is cost. A quantum calculation of the forces on each atom must be repeated at every step of a simulation, and the expense grows steeply with the number of atoms. A trained network answers almost instantly, so simulations can reach the larger cells and longer times at which melting, freezing and disorder actually show themselves. The network stands in for the calculation, not for the physics: its training data are still quantum results, and its purpose is to interpolate between them.
In the record
Limits
A network with no structural assumptions knows nothing about atoms: not that energy does not change when a crystal is rotated or shifted, nor that identical atoms are interchangeable. These must be supplied through hand-chosen descriptions of each atom's surroundings, and a poor choice limits what can be learned. Accuracy also holds only near the training data. Arrangements unlike anything it was shown, which is exactly what happens at a phase change or a defect, can be predicted confidently and wrongly, with no warning that the quantum calculation it replaced would have given a different answer.
Written by AIxSci: the background from general knowledge, the examples from checked records.
Related
Every article
aix-00276
Machine learning sorts a simulated polymer into ordered and disordered beads
aix-00272
Tree-based models predict sliding wear of carburised and plain engineering steels
aix-00220
Neural network predicts how titanium alloy grains round off during annealing
aix-00222
Neural networks trained on molecule pairs predict light absorption in stacks of fifty
aix-00216
Neural network models simulate how lead titanate loses its electrical polarisation on heating
aix-00213
Neural networks predict how much ultrasound energy drives magnesia dissolution
aix-00210
Language models suggest ingredients and firing temperatures for inorganic materials recipes
aix-00208
Neural network predicts atomic charges from bond lists, without atomic coordinates
aix-00206
Neural network designs five hard metallic glasses, all confirmed by experiment
aix-00205
Machine learning fitted to flexible molecules predicts oxidation potentials and hydration energies
aix-00160
Neural networks predict how zeolites take up carbon dioxide from structure alone
aix-00165
Thin molybdenum disulfide memory cells tested, with reinforcement learning fitting their circuit model
aix-00144
Machine-learned potentials trained to simulate copper ion conductor Cu7PS6 faster
aix-00133
Regression models predict how graphene loading changes aluminium's electron emission
aix-00131
Machine-learned potential predicts how chromium sulfide layers rearrange during exfoliation
aix-00119
Neural network potential used to simulate how calcium oxide melts under pressure
aix-00117
Deep learning model predicts polymer heat-softening across 48,208 designed candidates
aix-00113
Random atoms settle into molecules and crystals on a learned energy surface
aix-00108
Oxidation potentials for 15,238 molecules assembled from quantum calculations and lab measurements
aix-00103
Neural network predicts yield of chitosan nanoparticles grown with olive leaf extract
aix-00101
Simulations map three sodium–bismuth compounds, then models predict their heat-to-electricity performance
aix-00094
Neural network potential used to simulate carbon dioxide moving inside porous frameworks
aix-00091
Neural network force field simulates how a ferroelectric crystal's molecules begin to spin
aix-00088
Neural network retrained to predict how crystals absorb light, from few examples
aix-00071
Simulations map how surface atoms rearrange in alloys when molecules attach
aix-00070
Neural network scores candidate composite microstructures in a strength–toughness design loop
aix-00068
Machine learning predicts marine steel corrosion from seawater conditions and alloy make-up
aix-00058
Deep-learning model designs petrol blends from octane and soot targets
aix-00051
Ultrasound-treated clay bleaches cooking oil, with machine learning picking the settings
aix-00036
Four million simulated X-ray patterns used to test 21 crystal-symmetry classifiers
aix-00033
Neural network built from physics equations extracts solder deformation coefficients
aix-00028
Neural network trained on quantum calculations maps how gallium melts and freezes
aix-00027