Research

Each project is stated as a problem, my contribution, the result and its current status. Completed work and planned directions are labeled separately.

Preprint

Detectability and detection probability for model comparison

2025 – presentFirst author, project leadCOMPACT Collaboration · CWRU

What an expected information gain does, and does not, say about detecting a model in the one data set we have.

Problem
The Kullback–Leibler (KL) divergence is widely used to forecast whether a model can be distinguished from a null hypothesis. For two specified models it gives the expected log-likelihood ratio when data are drawn from the model in its first argument; it does not, by itself, tell us how often a search will detect a signal or recover its parameters.
Contribution
Conceived the project and developed a forecasting pipeline connecting statistical distinguishability to detection probabilities and parameter recovery. For multivariate Gaussian data, derived the distribution of the log-likelihood ratio and its cumulants under both hypotheses, and examined its links to discovery potential and parameter-estimation accuracy when the null is a point and the alternative a continuous family. Calibrated parameter searches with simulations to quantify false-positive rates, detection probabilities, and the effect of the detection criterion on recovered parameters.
Result
Detectability depends not only on the KL divergence but on the size of the parameter space, the likelihood-ratio distributions, and the distance from parameter-space boundaries (the look-elsewhere effect). The KL–detectability relation therefore has to be calibrated for each model family and test statistic. Applications to a toy Gaussian model, the cubic three-torus, the tensor-to-scalar ratio from B-mode polarization, and dark energy from supernova distances show why a forecast needs to account for the behavior of the analysis that will actually be performed.
Status
Preprint: arXiv:2610.17186

Ongoing

CMB statistics and covariance modeling for cosmic topology

2023 – presentProject leadCOMPACT Collaboration · CWRU

If space is finite, the CMB is no longer described by a power spectrum alone. This project computes what replaces it.

Problem
Non-trivial spatial topology breaks statistical isotropy, so the CMB requires full non-diagonal covariance matrices over (ℓ, m, ℓ′, m′) instead of an angular power spectrum.
Contribution
Derived the spin-2 (tensor) eigenmodes and CMB covariance matrices for orientable compact Euclidean manifolds, the first such computation, and showed that topology alone can induce parity-odd TB/EB correlations. Made a lead contribution to the eigenmodes, covariances and detectability of the non-orientable manifolds. Developed CMBtopology, the GPU-parallelized package used across the collaboration for tensor-mode likelihood analysis. In development: an end-to-end pipeline to constrain topology with Planck PR4 data, and, as a LiteBIRD adjoint member, a forecasting pipeline that uses E-mode polarization alongside temperature.
Result
Two first-author JCAP papers (Parts IIIa and IIIb) and a lead contribution as second author to a third (Part IIb). Over the multipole range computed, topological information in tensor-induced anisotropies keeps growing with ℓ rather than saturating as quickly as in the scalar case. A fourth paper (Part IIIc, first author) is in preparation.
Status
Ongoing: Part IIIa · Part IIIb · Part IIb · CMBtopology

In development

Neural likelihood emulation and simulation-based inference

CurrentProject leadCOMPACT Collaboration · CWRU

Learned inference that is checked: where does an approximation succeed, and where does it alter the inference?

Problem
Machine learning can make demanding analyses computationally feasible, but an accurate prediction of an observable does not by itself establish that the inferred constraints are reliable. A narrow posterior may reflect information in the data, restrictive simulation assumptions, or errors in a learned approximation. Bayesian parameter estimation for topology is also expensive, because searches must cover manifold size, orientation and observer position together.
Contribution
Leading work on neural likelihood emulators (variational autoencoders trained on simulated CMB data) and amortized inference, validated against conventional Bayesian methods. Because the likelihood can be evaluated directly in these problems, the comparison shows where learned approximations succeed and where they alter the inference; simulation-based calibration and coverage tests are part of the acceptance criterion. In a separate ongoing project with a student at the University of Toronto, investigating whether Koopman-operator methods can reduce the computational cost of diffusion-based posterior estimation without distorting the inferred distribution.
Result
No results yet. This work is in development.
Status
In development

Published

Learned classification benchmarked against a known likelihood

Dec 2023 – Mar 2024Co-authorCOMPACT Collaboration · CWRU / IFT Madrid

A machine-learning problem where the optimal answer is computable, so the network's shortfall can be measured.

Problem
Identifying the manifold behind a CMB map with the full likelihood is accurate but expensive.
Contribution
Contributed to benchmarking learned topology classification against a directly computable likelihood, and helped prepare and revise the manuscript.
Result
The paper shows that classifiers trained on simulated CMB maps approach likelihood-level discrimination at a fraction of the cost, and, because the likelihood is computable, measures the gap to the optimal decision rule.
Status
Published: JCAP 09 (2024) 057

Published · accepted

Calibrated evaluation and inference-time methods for language models

2025 – presentSecond author · co-authorCase Western Reserve University

The question behind the cosmology work, asked of a different system: how much evidence does one noisy measurement carry?

Problem
LLM benchmarks usually report a score without uncertainty, so rankings can be unstable and the inference procedure behind a reported number is rarely stated.
Contribution
Co-developed a Bayesian evaluation framework that replaces Pass@k with Dirichlet-posterior estimates and credible intervals, accounting for variation across repeated responses and for the uncertainty from a finite number of trials. Contributed to the evaluation and uncertainty methods of a framework for test-time scaling, and to the new version of the open-source Scorio toolkit.
Result
Don't Pass@k appeared at ICLR 2026 (main conference); the test-time-scaling paper is accepted at TMLR. An extension to reinforcement learning, where calibrated evaluation shapes the training signal, is underway.
Status
Published · accepted: ICLR 2026 · TMLR (accepted) · Scorio

Completed

Earlier research

Jan 2022 – Feb 2023Institute for Research in Fundamental Sciences (IPM), Tehran

  • Complex scalar-field dark matter in the non-relativistic limit: effective field theory, canonical transformations and multi-field axion models (advisor: M. H. Namjoo).
  • Cosmological perturbation theory, inflation and primordial non-Gaussianity using xAct (advisor: H. Firouzjahi).

Methods

Inference and statistics
Bayesian inference and MCMC · Neyman–Pearson likelihood-ratio testing · exact sampling distributions of test statistics · KL divergence and information-theoretic model comparison · ROC and coverage calibration · large-scale Monte Carlo design
Probabilistic machine learning
Uncertainty quantification and posterior calibration · Dirichlet-posterior evaluation · working knowledge of neural likelihood and likelihood-ratio estimation, VAEs as likelihood emulators, normalizing flows, score-based generative models
Deep learning
Transformers and LLMs (evaluation, uncertainty, test-time scaling) · PyTorch · JAX · scikit-learn · Ray
Computation
GPU / HPC pipelines · numerical linear algebra on large covariance matrices · Python · Julia · Mathematica · C
Cosmology
CMB temperature and polarization statistics · cosmic topology and eigenmode analysis · perturbation theory · primordial non-Gaussianity · Planck likelihood analysis

Collaborations

COMPACT
Cosmic topology collaboration: CWRU, Pittsburgh, Imperial College London, IFT Madrid, INFN Padova. Author lists are ordered first by contribution, then alphabetically.
LiteBIRD
CMB B-mode polarization satellite. Adjoint member, developing a forecasting pipeline to constrain cosmic topology using high-fidelity E-mode polarization alongside temperature.

Where this goes next: research directions, 2027 onward