An experiment-calibrated in-silico AI co-scientist discovers chemoselective silicon-nitride passivation for area-selective SiO2 atomic layer deposition
Year: 2026
Type: preprint
Abstract: Area-selective atomic layer deposition (AS-ALD) of SiO 2 -on-SiO 2 requires passivating adjacent SiN to isolate charge-trap nitride in 3D-NAND. We hypothesize that a smallmolecule electrophilic inhibitor chemoselectively acylates SiN surface amines (–NH 2 /–NH) over SiO 2 silanols (–OH), capping SiN against oxide nucleation. We build an in-silico AI co-scientist that (i) constructs amorphous a-SiO 2 /aSiN surfaces validated against diffraction and calibrates their reactive-site density to experiment, (ii) benchmarks its own methods, (iii) agentically screens inhibitors, and (iv) validates against a published positive control. The core selectivity is kinetic: reaction thermodynamics are near-degenerate, but acylation strongly favors the SiN amine over the SiO 2 silanol (gas-phase barrier difference ∼0.9 eV, a stepwise addition; xTB and DFT concur on the direction), reproducing textbook amine≫silanol chemose-lectivity. With a first-principles cap oxidative-initiation barrier (Ea≈0.68 eV ⇒ kdeg≈2.9× 10 −2 /cycle), a mechanistic multi-cycle model then demonstrates the target rather than assuming it: a re-passivation supercycle (redose every 3 cycles) reaches S = 0.91 at 10 nm. Run as a genuine closed loop, the agent autonomously proposes and screens candidates, inventing a novel lead that beats a naive electrophilicity-only baseline by 2.5 eV of selectivity. Our pipeline independently reproduces the published aminosilane oxide-preference (BDIPADS), anchoring the method. The pipeline is built entirely on open-source software.
Source: ChemRxiv
Author: Anand Sahu
Institution: Boston University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
Funder: Merck KGaA
OpenAlex ID: W7204933162