Neural Network Potentials Reveal Hidden Hydrogen Chemistry Inside Silicon Platelets

Hydrogen is the smallest atom in the universe, yet when it is blasted into a silicon wafer by the millions it can tear that crystal apart in ways that underpin one of the most important tricks in modern microelectronics: the transfer of ultrathin silicon layers. For decades, researchers have known that implanted hydrogen gathers into flat, disc-shaped defects known as platelets, which lie along the {111} crystallographic planes of silicon and act as the seeds from which blisters and exfoliation grow. What has remained stubbornly uncertain is exactly what the atoms inside and around these platelets look like, and how hydrogen arranges itself there. A new computational study published in the Journal of Materials Science by T. Yokoi, T. Hasegawa, Y. Ogura and K. Matsunaga of Nagoya University, in collaboration with the Japan Fine Ceramics Center, revisits this classic problem with a tool that simply did not exist in earlier eras of defect physics: a machine-learned neural-network potential capable of simulating systems far too large for conventional quantum-mechanical methods.

The central difficulty in studying hydrogen-induced platelets has always been one of scale. Density functional theory, the workhorse of first-principles materials simulation, can compute the energies and forces on atoms with high accuracy, but only for systems containing a few hundred atoms at most. A realistic platelet embedded in a silicon crystal, complete with edges, surrounding strain fields and a population of trapped hydrogen molecules, demands hundreds of thousands of atoms. Earlier theoretical pictures of platelet structure therefore relied on idealized, infinitely periodic models, in which the platelet extends endlessly in two dimensions and has no boundary. Such models capture the chemistry of the platelet interior but miss the physics of the edges, where the platelet meets pristine silicon and where much of the interesting hydrogen chemistry is expected to occur. The Nagoya team set out to close that gap.

Their approach centered on developing a neural-network potential, a class of interatomic potential in which a machine-learning model is trained on a large database of quantum-mechanical calculations and then learns to predict the energy and forces of any atomic configuration with near-DFT accuracy at a tiny fraction of the computational cost. The concept traces back to the high-dimensional neural-network framework introduced by Behler and Parrinello in 2007, and has since matured into a family of machine-learning potentials that are transforming computational materials science. The potential developed in this work was specifically constructed to be transferable to simulations involving hydrogen atoms in silicon, meaning that it remains accurate not just for one fixed bonding environment but across the wide variety of configurations hydrogen can adopt: bonded to silicon atoms at the platelet surfaces, sitting interstitially in the lattice, or existing as free H2 molecules within the platelet cavity.

A critical validation of the new potential was its ability to reproduce the relative stability of different hydrogen configurations in silicon. Hydrogen in crystalline silicon is famously versatile, occupying bond-centered and antibonding sites, forming Si-H bonds that passivate dangling bonds, pairing into molecules, and clustering into extended complexes. Previous experimental work, notably Raman scattering and infrared absorption studies dating back to the early 1990s, had identified the vibrational fingerprints of hydrogen molecules trapped inside {111}-oriented platelets, and theoretical studies had catalogued the energies of various hydrogen configurations. The neural-network potential was shown to predict these relative energies accurately, giving the researchers confidence that the large-scale simulations built on top of it would rest on a trustworthy quantum-mechanical foundation.

With the potential in hand, the team constructed simulation cells containing on the order of one hundred thousand atoms, large enough to embed a finite-sized platelet, with real edges, inside an otherwise perfect silicon lattice. This is the decisive methodological step. Because the platelet is finite, the simulation does not impose any artificial constraint on what happens inside it: the internal pressure exerted by hydrogen on the two facing platelet surfaces, and the resulting displacement of silicon atoms around the defect, emerge spontaneously from the molecular dynamics rather than being prescribed. The platelet can relax, its surfaces can separate or compress, and hydrogen atoms and molecules can migrate to whatever sites they find most favorable. In effect, the simulation lets the crystal solve its own structural problem.

The headline finding concerns hydrogen molecules. In the finite-sized platelets with low formation energies, the number of H2 molecules turned out to be two to three times greater than previous studies based on infinite-sized platelet models had predicted. This is not a small correction to a niche quantity. The population of molecular hydrogen inside a platelet is directly tied to the internal pressure that drives platelet growth, blistering and ultimately the exfoliation of thin silicon layers, the process exploited in silicon-on-insulator fabrication and layer-transfer technologies. If earlier periodic models systematically underestimated the molecular hydrogen content, then the quantitative picture of how platelets pressurize and crack open the crystal needs revision, and this study supplies the computational evidence for that revision.

Equally significant is where the molecules sit. The simulations revealed that H2 molecules reside at specific sites between the two platelet surfaces, rather than being distributed randomly through the cavity, and that this preferential placement contributes to lowering the formation energy of platelets that contain molecular hydrogen. In other words, the molecules are not passive passengers pumped up like a tire; they are structural participants whose positioning stabilizes the defect. This finding resonates with long-standing experimental observations of trapped molecular hydrogen in platelets, including Raman studies that detected the characteristic vibrational signature of H2 confined within {111}-oriented platelets, and it gives those spectroscopic observations a concrete atomistic interpretation.

The broader significance of the work lies in what it demonstrates about methodology. By directly modeling realistic finite-sized platelets with edges, the large-scale molecular simulations capture physics that periodic models structurally cannot: the elastic interaction between the platelet and the surrounding lattice, the role of the platelet periphery as a sink or source for hydrogen, and the spontaneous development of internal pressure. The authors argue that this capability provides deeper insight into hydrogen configurations and their energetics than was previously attainable. Given the growing family of machine-learning potentials, including Gaussian approximation potentials, moment tensor potentials and deep potential models, now being applied to hydrogenated silicon systems ranging from amorphous Si:H to heterointerfaces, the study positions neural-network potentials as the natural tool for revisiting other extended defects whose structures were frozen decades ago by the limits of first-principles cell sizes.

There is also a technological payoff. Hydrogen implantation and subsequent thermal treatment are used industrially to cleave off atomically thin silicon films, and the efficiency of that exfoliation depends sensitively on how quickly platelets nucleate, accumulate molecular hydrogen and coalesce into blisters. A corrected atomistic model of the platelet interior, one that accounts for the true H2 population and the stabilizing role of molecular hydrogen at specific inter-surface sites, feeds directly into more reliable predictions of blistering kinetics and layer-splitting conditions. The work, supported by the Japan Society for the Promotion of Science, the JST FOREST program and the Mitsubishi Foundation, with density functional calculations performed on the SQUID supercomputer at the University of Osaka, thus closes a loop that opened more than thirty years ago when transmission electron microscopists first imaged these tiny hydrogen-filled cracks. The smallest atom, it turns out, required one of the largest simulations, and a neural network, to finally show its hand.

Subject of Research: Atomic structure and energetics of hydrogen-induced {111} platelets in silicon modeled with a neural-network potential

Article Title: Revisiting atomic structures at hydrogen-induced platelets in silicon

Article References: Revisiting atomic structures at hydrogen-induced platelets in silicon. (n.d.). https://doi.org/10.1007/s10853-026-13902-6

Image Credits: AI Generated

DOI: 10.1007/s10853-026-13902-6

Keywords: silicon, hydrogen platelets, neural-network potential, molecular dynamics, H2 molecules, density functional theory, machine learning, defects, exfoliation, semiconductors, layer transfer, computational materials science