Aethorix v1.0: An Integrated Scientific AI Agent for Scalable Inorganic Materials Innovation and Industrial Implementation

TL;DR

The paper tackles the challenge of translating AI-driven materials design into industrial practice by introducing Aethorix v1.0, an integrated AI agent that combines diffusion-based structure generation, physics-informed surrogate models, and literature-grounded reasoning to enable inverse-design with ab initio fidelity at industrial timescales. It presents a closed-loop workflow that uses an LLM for problem framing, a diffusion-based generator for atomistic candidates, MLIPs for fast structure optimization, and physics-informed predictions to assess macroscale properties, all validated in a cement production use case. The cement study demonstrates problem framing, phase-stabilization insights in clinker systems, and adaptive process-parameter optimization via Bayesian methods, highlighting both the practical potential and current limitations (e.g., transferability across new dopants and kinetic effects). Overall, the work outlines a scalable path toward deploying AI across the materials-manufacturing value chain while acknowledging the need for continued improvement in transferability, multi-component handling, and dynamic control of complex processes.

Abstract

Artificial Intelligence (AI) is redefining the frontiers of scientific domains, ranging from drug discovery to meteorological modeling, yet its integration within industrial manufacturing remains nascent and fraught with operational challenges. To bridge this gap, we introduce Aethorix v1.0, an AI agent framework designed to overcome key industrial bottlenecks, demonstrating state-of-the-art performance in materials design innovation and process parameter optimization. Our tool is built upon three pillars: a scientific corpus reasoning engine that streamlines knowledge retrieval and validation, a diffusion-based generative model for zero-shot inverse design, and specialized interatomic potentials that enable faster screening with ab initio fidelity. We demonstrate Aethorix's utility through a real-world cement production case study, confirming its capacity for integration into industrial workflows and its role in revolutionizing the design-make-test-analyze loop while ensuring rigorous manufacturing standards are met.

Figures (6)

• Figure 1: Schematic Illustration of the Aethorix v1.0 Workflow and Its Industrial Applications. Aethorix v1.0 integrates LLM-guided problem definition, generative structure design, structure optimization, and physics-informed prediction into an iterative loop for solving industrial challenges. Candidate solutions are refined until validated under real-world conditions, enabling prototype testing and industrial deployment. The scientific AI agent underpins applications including but not limited to novel material formulations design, tailored material design improvement, process protocol optimization, and lifetime reliability forecasting.
• Figure 2: Computational Cost Comparison for Common Materials Emergent Properties. Hollow and solid markers denote tasks completed using DFT and MLIP, respectively. MLIP shows 3–5 orders of magnitude greater computational economy compared to DFT for deriving emergent properties. Costs of DFT calculations and MLIP inferences were benchmarked on an Intel Xeon Platinum 8175M CPU (32 cores) and on a single NVIDIA Tesla T4 GPU, respectively. All hardware instances were provisioned through Amazon Web Services.
• Figure 3: Translating Problem Statements into Design Principles.a, Knowledge synthesis of scientific literature by the integrated SciSpace module, identifying problem root causes through evidence-based reasoning and proposing solution strategies based on the agent's capability. b, Solution formulation facilitated by the Aethorix v1.0 agent, where user interaction enables the development and iterative refinement of an actionable implementation plan. The agent's responses (blue) and user's feedback (orange) are color-coded.
• Figure 4: Trace Element Stabilization of C$_3$S Clinker Phases.a, Elemental scope of the customized AethIP potential. The base model was trained on the Ca-Mg-Si-O system of the host structure, while extended models were fine-tuned to include specific dopant elements. b, Active learning history of the base AethIP model. Model transferability was assessed on-the-fly by quantifying the number of uncertain structures within newly generated ensembles. A structure was flagged as uncertain if the force standard deviation $>$ 1 eV/Å and/or energy standard deviation $>$ 40 meV/atom. c, Transferability test of base AethIP on three unseen C$_3$S and C$_2$S polymorphs, respectively. AethIP significantly outperforms both UMA-M-OC20 and MACE-MPA-0 for structural optimization and formation energy prediction. d, Transferability test of extended AethIP on six doped C$_3$S single phase. AethIP exceeds the accuracy of MACE-MPA-0 but not that of UMA-M-OC20. e, Gibbs free energies of formation as a function of temperature (0–2000 K) for the 50 most stable C$_3$S polymorphic phases identified at 0 K. f, Ionic self-diffusivities at 300 K for the ten most stable C$_3$S polymorphs synthesized at 1730 K. Lattice structures are visualized above the corresponding bars, ordered by descending thermodynamic stability at 1730 K. g, Heat map of the formation Gibbs free energy change induced by 10 mol% dopant substitution at each crystallographic site in the most stable C$_3$S polymorphs synthesized at 1730 K. The color scale represents the formation energy difference relative to the bare (undoped) structure at 0 K.
• Figure 5: Enhanced Cement Compressive Strength Prediction via Multimodal ML. Performance comparision between empirical, purely data-driven and multimodal ML methods, evaluated through a, root mean squared error (RMSE), b, coefficient of determination ($R^2$), and c, prediction accuracy within thresholds of $\pm 1.5$ MPa and $\pm 1.0$ MPa. d, Optimized parameter deviations from operational means for compressive strength maximization across curing periods, expressed as percentage differences. Deviations less than 1% are not shown on the plot.