An Amdahl Framework for Agentic Computational Science
- 1Department of Chemical and Biomolecular Engineering
- 2KAIST
Abstract
AI agents increasingly compress human-mediated parts of computational research, shifting bottlenecks toward whatever remains slow. Amdahl-like models capture this serial-bottleneck logic but usually take the non-compressible block as fixed. Here I ask what that block contains in computational science. I start from a fixed-workflow baseline separating numerical execution and resource waiting from active human work and synchronization delay, then relax the assumption that nominal simulation time is irreducible. Human researchers already provide semantic oversight intermittently; persistent agents can apply the same scrutiny earlier and more consistently. This motivates a counterfactual split between core and avoidable computation, and a closed-loop ceiling in which only minimum necessary critical-path computation remains irreducible after accounting for validation cost. As human-mediated delays shrink, bottlenecks migrate toward numerical kernels and resource constraints, making faster solvers, surrogate models and high-performance computing more valuable.
Related
- Science Done on a Machine by a Machine: AI Agents in Computational Chemistry
- Automated Many-Body Simulations of Strongly Correlated Systems Using a Correlation-Aware Agentic Framework
- A traceable agentic framework for chemistry research query and analysis with ThermoML database
- OptoMatKG: A Traceable Knowledge Graph Framework with Semantic Retrieval for Optoelectronic Materials Science
- Thermodynamic regularization based computational framework for strain-induced crystallization
BibTeX
@misc{kim2026amdahl,
title = {{An Amdahl Framework for Agentic Computational Science}},
author = {Jihan Kim},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008151/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008151/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - An Amdahl Framework for Agentic Computational Science
AU - Jihan Kim
PY - 2026
DA - 2026/09/01
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008151/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008151/v1
AB - AI agents increasingly compress human-mediated parts of computational research, shifting bottlenecks toward whatever remains slow. Amdahl-like models capture this serial-bottleneck logic but usually take the non-compressible block as fixed. Here I ask what that block contains in computational science. I start from a fixed-workflow baseline separating numerical execution and resource waiting from active human work and synchronization delay, then relax the assumption that nominal simulation time is irreducible. Human researchers already provide semantic oversight intermittently; persistent agents can apply the same scrutiny earlier and more consistently. This motivates a counterfactual split between core and avoidable computation, and a closed-loop ceiling in which only minimum necessary critical-path computation remains irreducible after accounting for validation cost. As human-mediated delays shrink, bottlenecks migrate toward numerical kernels and resource constraints, making faster solvers, surrogate models and high-performance computing more valuable.
ER -