Title:	Computational Organic Energy Materials Design (COEMD)
DNr:	NAISS 2025/3-65
Project Type:	NAISS Large
Principal Investigator:	Carlos Moyses Graca Araujo <moyses.araujo@kau.se>
Affiliation:	Karlstads universitet
Duration:	2026-01-01 – 2026-07-01
Classification:	10304 10402 10105
Keywords:

Abstract

Organic energy materials combine sustainability with exceptional chemical tunability, but the mechanisms linking atomic structure, mesoscale morphology, and device performance remain only partly understood. We propose an integrated computational program to accelerate discovery in three tightly connected domains: (i) polymer:non-fullerene acceptor (NFA) blends for organic photovoltaics (OPVs), (ii) redox-active organic electrodes for post-Li batteries, and (iii) organic photocatalysts for solar-driven hydrogen production. Building on our validated first-stage workflow for OPVs -- where molecular dynamics (MD) coupled to density-functional theory (DFT) reproduced the electronic and optical signatures of PF5Y5 -- we will extend to representative NFAs (Y5/Y6 families, perylene and indacenodithiophene derivatives) and efficient polymer donors (e.g., PBDB-T), while in parallel designing programmable redox backbones for organic batteries (quinone, TEMPO, triazine scaffolds; linear ionomers, cross-linked networks, porous CMP/COF polymers, polymer–nanocarbon hybrids) and conjugated organic photocatalysts for H₂ evolution. The core methodology is a multiscale MD → sequential QM/MM DFT/TDDFT pipeline: solvent-aware MD (including evaporation protocols and replica sampling) provides realistic morphologies and local environments; S-QM/MM computes redox thermodynamics, excited states, level alignment, transport proxies, and interfacial kinetics across statistically meaningful ensembles. To accelerate screening and close the design loop, we will train neural surrogates (graph neural networks and Transformers) on MD/DFT/TDDFT descriptors, and deploy generative models (VAE/flow/diffusion) to propose chemically valid candidates conditioned on performance. Active-learning strategies with uncertainty quantification will iteratively select structures for high-fidelity recalculation, maximizing information gained per compute hour. The project requires high-performance computing to (i) simulate systems of 50–200k atoms for hundreds of nanoseconds to capture morphology and disorder, (ii) evaluate hundreds to thousands of QM/QM-embedded excited-state calculations per material family, and (iii) train large neural models on long molecular sequences and graphs. We will use Tetralith@NSC and Dardel@PDC for CPU-intensive MD and S-QM/MM (including fat-memory nodes for TDDFT), and Alvis GPUs for neural-network training, hyper-parameter search, inference, and active learning. Efficient job sizing (32–48 CPU cores for 50–60k-atom MD; mixed-precision training on A100 GPUs) and lightweight-FAIR data practices (versioned models, scripted preprocessing, compact derived features) ensure high utilization and reproducibility. Expected outcomes include validated structure–property maps under realistic processing histories, actionable design rules for programmable redox backbones and continuous electronic/ionic pathways at high areal loading, down-selected candidates for experimental verification in OPVs, organic batteries, and organic photocatalysis, and openly shared datasets and software components. By combining high-fidelity physics with scalable AI on the Swedish National Academic Infrastructure for Supercomputing (NAISS), the project will advance fundamental understanding and deliver practical leads for sustainable energy technologies.