Research

Publications

Computation, machine learning, and AI applied across the sciences and beyond, from molecular simulation and statistical thermodynamics to multi-agent AI systems. Full list on Google Scholar.

Essays Less formal writing on AI, science, and building, on Substack →
  1. Figure from “What LLM Agents Say When No One Is Watching: Social Structure and Latent Objective Emergence in Multi-Agent Debates”
    A dual-channel (public vs. off-the-record) probe reveals latent objectives emerging in multi-agent LLM debates.

    What LLM Agents Say When No One Is Watching: Social Structure and Latent Objective Emergence in Multi-Agent Debates

    S. A. Ghaffarizadeh, D. Mohaddes, A. Izadkhah, S. Noroozizadeh

    arXiv (preprint) · 2026

    Introduces a dual-channel test (pairing each agent's public statement with a private, off-the-record one) to surface what LLM agents in a multi-agent debate actually intend. Across ten models, social pressure to conform drove a targeted agent's public and private decisions apart, from a roughly 3% baseline to about 40%, arguing that agent evaluation must look past stated goals to catch emergent objectives.

  2. Figure from “A Picture is Worth a Thousand Timesteps: Excess Entropy Scaling for Rapid Estimation of Diffusion Coefficients in Molecular-Dynamics Simulations of Fluids”
    Excess-entropy scaling for rapid estimation of diffusion coefficients from short MD trajectories.

    A Picture is Worth a Thousand Timesteps: Excess Entropy Scaling for Rapid Estimation of Diffusion Coefficients in Molecular-Dynamics Simulations of Fluids

    S. A. Ghaffarizadeh, G. J. Wang

    Journal of Chemical Theory and Computation · 2024

    Shows that excess-entropy scaling estimates diffusion coefficients from short molecular-dynamics runs far more accurately than the standard Einstein–Helfand or Green–Kubo methods at equal sampling time. Because it needs only structural information (the radial distribution function), a single structural snapshot can stand in for long trajectories.

  3. Figure from “The environmental footprint of bitcoin mining across the globe: Call for urgent action”
    Global water, land, and carbon footprint of Bitcoin mining.

    The environmental footprint of bitcoin mining across the globe: Call for urgent action

    S. Chamanara, S. A. Ghaffarizadeh, K. Madani

    Earth's Future · 2023

    A multi-attribute accounting of Bitcoin mining's global environmental toll over 2020 to 2021: 173 TWh of electricity, 86 Mt of CO₂-equivalent emissions, and a 1.65 km³ water footprint. By quantifying land, water, and carbon costs together, it makes the case for urgent action from the scientific, policy, and advocacy communities.

  4. Figure from “Getting over the hump with KAMEL-LOBE: Kernel-averaging method to eliminate length-of-bin effects in radial distribution functions”
    KAMEL-LOBE removes length-of-bin artifacts in radial distribution functions.

    Getting over the hump with KAMEL-LOBE: Kernel-averaging method to eliminate length-of-bin effects in radial distribution functions

    S. A. Ghaffarizadeh, G. J. Wang

    The Journal of Chemical Physics · 2023

    Demonstrates that the arbitrary bin width used to histogram radial distribution functions can introduce spurious artifacts into common analyses such as phase-boundary detection and excess-entropy scaling. It introduces KAMEL-LOBE, a mass-conserving Gaussian-kernel smoothing that cleanly removes these length-of-bin effects.

  5. Figure from “Excess entropy scaling in active-matter systems”
    Excess-entropy scaling collapses transport coefficients in active-matter systems.

    Excess entropy scaling in active-matter systems

    S. A. Ghaffarizadeh, G. J. Wang

    The Journal of Physical Chemistry Letters · 2022

    Uses molecular-dynamics simulations of active matter (particles that self-propel by consuming free energy) to establish an excess-entropy scaling relation in a driven, non-equilibrium system. The result connects transport dynamics to static structure, extending a tool previously confined to equilibrium fluids.

  6. Figure from “Ultrahigh evaporative heat transfer measured locally in submicron water films”
    Deep Neural Network prediction of evaporative heat transfer in submicron water films.

    Ultrahigh evaporative heat transfer measured locally in submicron water films

    X. Wang, S. A. Ghaffarizadeh, A. J. H. McGaughey, J. A. Malen

    Scientific Reports · 2022

    Resolves the evaporation heat-transfer coefficient locally within a submicron water film by pairing non-contact thermoreflectance measurements with a neural-network surrogate trained on finite-element simulations. It reveals ultrahigh interfacial evaporation, two orders of magnitude above conventional values, relevant to cooling high-power micro/nano-devices.

  7. Figure from “Fishers Handle Bugs Better than Fish-Receivers: Nourishing Computational Self-Efficacy in Engineering Coursework”
    Students who write code debug more effectively than those who only receive it.

    Fishers Handle Bugs Better than Fish-Receivers: Nourishing Computational Self-Efficacy in Engineering Coursework

    S. A. Ghaffarizadeh, G. J. Wang

    ASEE Annual Conference & Exposition · 2022

    Studies how to build computational self-efficacy in a graduate engineering course where students incrementally write a full molecular-simulation code. It finds that allowing assignment resubmission measurably improves outcomes, most of all for students with the least prior programming exposure, supporting a growth-mindset approach to teaching code.

  8. Figure from “Life and work of researchers trapped in the COVID-19 pandemic vicious cycle”
    How the pandemic reshaped researchers' work and productivity.

    Life and work of researchers trapped in the COVID-19 pandemic vicious cycle

    S. A. Ghaffarizadeh, et al.

    bioRxiv (preprint) · 2021

    Surveys 740 researchers on how the COVID-19 pandemic reshaped their work, finding two-thirds lost productivity and felt heightened pressure to make progress. Disruption fell hardest on lab- and field-based work and on caregivers, with a pronounced gender gap, offering evidence to guide institutional and policy responses.

  9. Figure from “Specific surface area from nitrogen adsorption data at 77 K using the zeta adsorption isotherm”
    Specific surface area from N₂ adsorption via the zeta adsorption isotherm.

    Specific surface area from nitrogen adsorption data at 77 K using the zeta adsorption isotherm

    S. A. Ghaffarizadeh, S. H. Zandavi, C. A. Ward

    The Journal of Physical Chemistry C · 2017

    Proposes a method to determine the specific surface area of solid powders from nitrogen-adsorption measurements at 77 K using the zeta adsorption isotherm. Validated across six powders spanning two orders of magnitude in surface area, it yields consistent values with uncertainty recoverable from a single equilibrium measurement.