TritonDFT: Automating DFT with a Multi-Agent Framework

Zhengding Hu; Kuntal Talit; Zhen Wang; Haseeb Ahmad; Yichen Lin; Prabhleen Kaur; Christopher Lane; Elizabeth A. Peterson

cond-mat.mtrl-sciMar 2, 2026

TritonDFT: Automating DFT with a Multi-Agent Framework

Authors: Zhengding Hu, Kuntal Talit, Zhen Wang, Haseeb Ahmad, Yichen Lin, Prabhleen Kaur, Christopher Lane, Elizabeth A. Peterson,

Organizations: University of California, San Diego · University of California, Merced · Los Alamos National Laboratory

Abstract

Density Functional Theory (DFT) is a cornerstone of materials science, yet executing DFT in practice requires coordinating a complex, multi-step workflow. Existing tools and LLM-based solutions automate parts of the steps, but lack support for full workflow automation, diverse task adaptation, and accuracy-cost trade-off optimization in DFT configuration. To this end, we present TritonDFT, a multi-agent framework that enables efficient and accurate DFT execution through an expert-curated, extensible workflow design, Pareto-aware parameter inference, and multi-source knowledge augmentation. We further introduce DFTBench, a benchmark for evaluating the agent's multi-dimensional capabilities, spanning science expertise, trade0off optimization, HPC knowledge, and cost efficiency. TritonDFT provides an open user interface for real-world usage. Our website is at https://www.tritondft.com. Our source code and benchmark suite are available at https://github.com/Leo9660/TritonDFT.git.

Figures & tables

Figure 1: DFT execution is a complex, multi-step process requiring heterogeneous domain expertise. Based on an internal survey conducted with 19 domain researchers at the PhD level or above, manually handling each step typically takes minutes to hours. TritonDFT reduces the per-step time to the scale of seconds to minutes, and provides automation across the entire workflow.

Figure 2: Energy Deviation and Computational Cost Variations with different DFT Parameters for Silicon (Space Group

Fd\bar{3}m

). Computational cost is normalized as

C/C_{0}

, where

C

denotes the actual execution time and

C_{0}

corresponds to the execution time with K=6 and ecut=40.

Method	Framework Architecture			Evaluation Dataset & Metrics
	Supported Task Types	Parameter Configuration	Knowledge Augmentation	Number of Material Types	Ground Truth Curation	Accuracy-Cost Tradeoff	Parallel Efficiency	Monetary Cost
DREAMS ( Wang et al., 2025 )	Surface Chemistry (Adsorption)	Physics Only	Open Database	2 (Metal, Insulator)	Public Dataset	✗	✗	✗
VASPilot ( Liu et al., 2025 )	Electronic Structure (Band, DOS)	Physics Only	Open Database	1 (Semiconductor)	Public Dataset	✗	✗	✗
AgenticDFT ( Xia et al., 2025 )	Geometry & Energetics (Relaxation, Band)	Physics Only	Open Database	2 (Metal, Semiconductor)	Public Dataset	✗	✗	✗
TritonDFT (Ours)	General QE Usage ( $>10$ Types)	Physics + HPC (Pareto-aware)	Open Database + Memory + Human Interact.	10 (Metal, Insulator, van der Waals, Topological, …)	Expert Curated Calculation	✓	✓	✓

Table 1: Comparison of TritonDFT with state-of-the-art agentic DFT frameworks.

Figure 3: The overview of TritonDFT framework.

Model	$\Delta E<$ 20 meV/atom		$\Delta E<$ 10 meV/atom		$\Delta E<$ 1 meV/atom		Advanced Parameter Satisfaction Rate
Model	Pass Rate	Norm. Cost	Pass Rate	Norm. Cost	Pass Rate	Norm. Cost	Advanced Parameter Satisfaction Rate
GPT 5.2	70.5%	14.29	67.0%	8.95	47.1%	4.23	51.3%
GPT 5.1	39.3%	6.22	32.9%	4.21	9.8%	2.24	43.6%
GPT 4o	52.8%	1.85	38.2%	1.28	13.6%	0.50	28.2%
GPT 4o mini	5.7%	1.01	5.6%	1.17	4.5%	0.97	28.2%
Gemini 2.5 Pro	59.6%	3.77	53.9%	2.95	14.9%	1.24	48.7%

Table 2: Model performance on DFT parameter configuration across different LLMs under varying error threshold.

Figure 4: Performance Analysis with Pass Rate and Cost Efficiency across different material types. Cost Efficiency is measured (1 / Cost Factor), averaged over all passed cases within each type.

Model	VC-relax (Lattice)	SCF (E/atom)	Bandgap (eV)	E F (eV)
GPT 5.2	5.7%	0.0%	0.34	1.32
GPT 5.1	7.4%	0.1%	0.45	2.34
GPT 4o	10.1%	1.2%	1.90	8.17
Gemini 2.5 Pro	5.5%	0.1%	1.02	1.45
Gemini 2.5 Flash	8.1%	0.5%	1.19	7.09
Claude Opus 4.5	8.0%	0.1%	1.84	3.01

Table 3: Mean relative error (MRE, %) for VC-relax and SCF, and mean absolute error (MAE) for Band Gap and DOS, computed over successfully finished execution results.

Model	16 Cores	32 Cores	64 Cores
GPT 5.2	+14.4%	+4.2%	+15.1%
GPT 5.1	+11.3%	-5.8%	-14.1%
GPT 4o	-21.0%	-34.0%	-23.6%
GPT 4o mini	-25.7%	-25.6%	+2.8%
Gemini 2.5 Pro	4.74%	-6.4%	-3.29%
Gemini 2.5 Flash	-20.7%	-43.7%	-32.0%

Table 4: Relative speedup (

\%

) of parallel execution of different LLMs over the default baseline across different CPU core numbers.

Figure 5: Comparison of TritonDFT ’s workflow throughput with different LLMs across different DFT tasks.

Model	Struct Relax	Band Gap	DOS
GPT 5.2	0.05 $\pm$ 0.02	0.15 $\pm$ 0.04	0.13 $\pm$ 0.04
GPT 5.1	0.04 $\pm$ 0.02	0.13 $\pm$ 0.04	0.10 $\pm$ 0.03
GPT 4o	0.06 $\pm$ 0.03	0.18 $\pm$ 0.04	0.14 $\pm$ 0.03
Gemini 2.5 Pro	0.05 $\pm$ 0.03	0.13 $\pm$ 0.04	0.11 $\pm$ 0.04
Gemini 2.5 Flash	0.01 $\pm$ 0.01	0.03 $\pm$ 0.01	0.03 $\pm$ 0.01
Claude Opus 4.5	0.15 $\pm$ 0.08	0.44 $\pm$ 0.13	0.37 $\pm$ 0.11

Table 5: Average cost consumption (USD) per query regarding API usage across different tasks.

Figure 6: Accuracy-Cost Trade-offs across different models and parameter inference methods when using One-Shot inference and Pareto-aware inference, respectively.

Figure 7: Comparison of TritonDFT ’s results with human-curated results on phonon dispersion calculation.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

#	Material	Category	Space Group	Atoms
1	CaO	Insulator	Fm $\bar{3}$ m	2
2	KCl	Insulator	Fm $\bar{3}$ m	2
3	LiF	Insulator	Fm $\bar{3}$ m	2
4	MgF 2	Insulator	P4 2 /mnm	6
5	MgO	Insulator	Fm $\bar{3}$ m	2
6	NaCl	Insulator	Fm $\bar{3}$ m	2

Appendix

Table 6: Benchmark dataset of materials used for DFT workflow evaluation, including material category (multiple categories separated by commas), space group, and number of atoms in the conventional unit cell.

Scientific Workflow Automation

Scientific workflow automation uses software, AI agents, and laboratory systems to plan, execute, validate, and reproduce research tasks.

See all 160 papers

AutoDFT: A Closed-Loop Multi-Agent Framework for Autonomous DFT Calculations
May 25, 2026Penghui Yang, Zhonghan Zhang, Yue Li +6Multi-Agent LLM Systems Scientific Workflow Automation
Nanyang Technological University, Singapore · Singapore Management University
AGAPI-Agents: An Open-Access Agentic AI Platform for Accelerated Materials Design on AtomGPT.org
Dec 12, 2025Jaehyung Lee, Justin Ely, Kent Zhang +3AI Agents for Scientific Discovery LLM Agent Evaluation
Department of Materials Science and Engineering, Johns Hopkins University, Baltimore, MD 21218, USA · Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218, USA · Department of Physics and Astronomy, West Virginia University, Morgantown, WV 26506, USA
A fully differentiable framework for training proxy Exchange Correlation Functionals for periodic systems
Feb 17, 2026Rakshit Kumar Singh, Aryan Amit Barsainyan, Bharath RamsundarComputational Quantum Chemistry Density Functional Theory
Deep Forest Sciences

May 25, 2026J/K move · Enter open · S save

AutoDFT: A Closed-Loop Multi-Agent Framework for Autonomous DFT Calculations

Penghui Yang, Zhonghan Zhang, Yue Li, Xinrun Wang+5

Nanyang Technological University, Singapore · Singapore Management University

Multi-Agent LLM Systems Scientific Workflow Automation LLM Agent Planning Computational Materials Science

Open paper AskPDF arXiv

Density functional theory (DFT) serves as the basis for computational discovery in materials science and chemistry, yet each calculation demands extensive human effort: adjusting algorithms when convergence stalls, revising plans when unexpected physics emerges, and inserting steps as intermediate results reshape the problem. Existing LLM-based agents automate only the initial planning stage, producing a full execution plan upfront and leaving all subsequent adaptation to hand-crafted rules. As a result, these workflows remain fragile, do not generalize well beyond pre-planned scenarios, and often require expert intervention when failures or unexpected intermediate results require changes to the calculation path. Here, we introduce AutoDFT, a closed-loop multi-agent framework that embeds LLM reasoning into every stage of the DFT lifecycle, where a strategic planner produces a skeletal plan of step objectives; a step planner generates numerical parameters just in time from preceding results; and a monitor-recover-reflect cycle diagnoses failures, repairs them, and revises the plan when the evidence justifies it. We demonstrate both breadth and depth: breadth on VASPBench, a purpose-built benchmark spanning 34 tasks and 9 DFT calculation types, where AutoDFT achieves 94.1% task-level success with GPT-5.2; and depth on established materials databases, where AutoDFT produces quantitatively reliable property predictions across electronic, magnetic, and energetic properties. By closing the loop between planning and execution, AutoDFT enables experimentalists without deep computational expertise to obtain reliable first-principles results.

Abstract

Scientific Workflow Automation

Explore similar work

AutoDFT: A Closed-Loop Multi-Agent Framework for Autonomous DFT Calculations

AGAPI-Agents: An Open-Access Agentic AI Platform for Accelerated Materials Design on AtomGPT.org

A fully differentiable framework for training proxy Exchange Correlation Functionals for periodic systems