Back within the old days — the really old days — the duty of designing materials was laborious. Investigators, over the course of 1,000-plus years, tried to make gold by combining things like lead, mercury, and sulfur, mixed in what they hoped could be just the proper proportions. Even famous scientists like Tycho Brahe, Robert Boyle, and Isaac Newton tried their hands on the fruitless endeavor we call alchemy.
Materials science has, in fact, come a good distance. For the past 150 years, researchers have had the advantage of the periodic table of elements to attract upon, which tells them that different elements have different properties, and one can’t magically transform into one other. Furthermore, up to now decade or so, machine learning tools have considerably boosted our capability to find out the structure and physical properties of varied molecules and substances. Latest research by a gaggle led by Ju Li — the Tokyo Electric Power Company Professor of Nuclear Engineering at MIT and professor of materials science and engineering — offers the promise of a significant leap in capabilities that may facilitate materials design. The outcomes of their investigation are reported in a December 2024 issue of
At present, many of the machine-learning models which are used to characterize molecular systems are based on density functional theory (DFT), which offers a quantum mechanical approach to determining the whole energy of a molecule or crystal by taking a look at the electron density distribution — which is, principally, the common variety of electrons situated in a unit volume around each given point in space near the molecule. (Walter Kohn, who co-invented this theory 60 years ago, received a Nobel Prize in Chemistry for it in 1998.) While the strategy has been very successful, it has some drawbacks, in accordance with Li: “First, the accuracy is just not uniformly great. And, second, it only tells you one thing: the bottom total energy of the molecular system.”
“Couples therapy” to the rescue
His team is now counting on a distinct computational chemistry technique, also derived from quantum mechanics, generally known as coupled-cluster theory, or CCSD(T). “That is the gold standard of quantum chemistry,” Li comments. The outcomes of CCSD(T) calculations are far more accurate than what you get from DFT calculations, they usually will be as trustworthy as those currently obtainable from experiments. The issue is that carrying out these calculations on a pc may be very slow, he says, “and the scaling is bad: When you double the variety of electrons within the system, the computations turn out to be 100 times dearer.” For that reason, CCSD(T) calculations have normally been limited to molecules with a small variety of atoms — on the order of about 10. Anything much beyond that will simply take too long.
That’s where machine learning is available in. CCSD(T) calculations are first performed on conventional computers, and the outcomes are then used to coach a neural network with a novel architecture specially devised by Li and his colleagues. After training, the neural network can perform these same calculations much faster by profiting from approximation techniques. What’s more, their neural network model can extract far more details about a molecule than simply its energy. “In previous work, people have used multiple different models to evaluate different properties,” says Hao Tang, an MIT PhD student in materials science and engineering. “Here we use only one model to judge all of those properties, which is why we call it a ‘multi-task’ approach.”
The “Multi-task Electronic Hamiltonian network,” or MEHnet, sheds light on quite a few electronic properties, similar to the dipole and quadrupole moments, electronic polarizability, and the optical excitation gap — the quantity of energy needed to take an electron from the bottom state to the bottom excited state. “The excitation gap affects the optical properties of materials,” Tang explains, “since it determines the frequency of sunshine that will be absorbed by a molecule.” One other advantage of their CCSD-trained model is that it could actually reveal properties of not only ground states, but additionally excited states. The model can even predict the infrared absorption spectrum of a molecule related to its vibrational properties, where the vibrations of atoms inside a molecule are coupled to one another, leading to varied collective behaviors.
The strength of their approach owes quite a bit to the network architecture. Drawing on the work of MIT Assistant Professor Tess Smidt, the team is utilizing a so-called E(3)-equivariant graph neural network, says Tang, “by which the nodes represent atoms and the sides that connect the nodes represent the bonds between atoms. We also use customized algorithms that incorporate physics principles — related to how people calculate molecular properties in quantum mechanics — directly into our model.”
Testing, 1, 2 3
When tested on its evaluation of known hydrocarbon molecules, the model of Li et al. outperformed DFT counterparts and closely matched experimental results taken from the published literature.
Qiang Zhu — a materials discovery specialist on the University of North Carolina at Charlotte (who was not a part of this study) — is impressed by what’s been completed to this point. “Their method enables effective training with a small dataset, while achieving superior accuracy and computational efficiency in comparison with existing models,” he says. “That is exciting work that illustrates the powerful synergy between computational chemistry and deep learning, offering fresh ideas for developing more accurate and scalable electronic structure methods.”
The MIT-based group applied their model first to small, nonmetallic elements — hydrogen, carbon, nitrogen, oxygen, and fluorine, from which organic compounds will be made — and has since moved on to examining heavier elements: silicon, phosphorus, sulfur, chlorine, and even platinum. After being trained on small molecules, the model will be generalized to greater and larger molecules. “Previously, most calculations were limited to analyzing a whole bunch of atoms with DFT and just tens of atoms with CCSD(T) calculations,” Li says. “Now we’re talking about handling hundreds of atoms and, eventually, perhaps tens of hundreds.”
For now, the researchers are still evaluating known molecules, however the model will be used to characterize molecules that haven’t been seen before, in addition to to predict the properties of hypothetical materials that consist of various sorts of molecules. “The concept is to make use of our theoretical tools to select promising candidates, which satisfy a specific set of criteria, before suggesting them to an experimentalist to ascertain out,” Tang says.
It’s all in regards to the apps
Looking ahead, Zhu is optimistic in regards to the possible applications. “This approach holds the potential for high-throughput molecular screening,” he says. “That’s a task where achieving chemical accuracy will be essential for identifying novel molecules and materials with desirable properties.”
Once they exhibit the flexibility to research large molecules with perhaps tens of hundreds of atoms, Li says, “we must always have the opportunity to invent recent polymers or materials” that is perhaps utilized in drug design or in semiconductor devices. The examination of heavier transition metal elements may lead to the appearance of recent materials for batteries — presently an area of acute need.
The longer term, as Li sees it, is wide open. “It’s not about only one area,” he says. “Our ambition, ultimately, is to cover the entire periodic table with CCSD(T)-level accuracy, but at lower computational cost than DFT. This could enable us to resolve a big selection of problems in chemistry, biology, and materials science. It’s hard to know, at present, just how wide that range is perhaps.”
This work was supported by the Honda Research Institute. Hao Tang acknowledges support from the Mathworks Engineering Fellowship. The calculations on this work were performed, partly, on the Matlantis high-speed universal atomistic simulator, the Texas Advanced Computing Center, the MIT SuperCloud, and the National Energy Research Scientific Computing.
