How soft metal halide perovskite lattices heal following energetic particle bombardment
First authorDOI
Machine Learning Engineer · Johns Hopkins · LLNL
Kumar Miskin
PhD candidate in Materials Science & Engineering
Building physics-informed ML systems that accelerate scientific discovery — graph neural networks, Bayesian optimization, and active learning for materials, energy, and space applications.
4
Peer-reviewed publications
44
Citations
2
First-author papers
3
Continents, industry + national lab
01 · About
Bridging ML and materials
I'm a Machine Learning Engineer and PhD candidate in Materials Science & Engineering at Johns Hopkins University, advised by Prof. Paulette Clancy, graduating December 2026. My work sits at the intersection of machine learning and physical simulation — building systems that learn from atomistic data to accelerate scientific discovery.
My research has taken me from a U.S. DOE national laboratory (Lawrence Livermore) to global industry (Sony, Tata Power) and academic groups across three continents. Along the way I've shipped production ML systems for computer vision, audio, and autonomous material discovery — and co-authored four peer-reviewed papers on physics-informed ML and materials.
Beyond the lab, I mentor students, teach graduate courses, and build full-stack AI products end-to-end.
Focus areas
Machine learning interatomic potentialsGraph neural networksBayesian optimizationActive learningHalide perovskitesSolid-state electrolytes & hydroboratesDefect & radiation physicsComputational materials discovery
Education
Ph.D. in Materials Science & Engineering
Johns Hopkins University
M.S. 2024 · Advisor: Prof. Paulette Clancy · Dissertation: physics-informed ML for scientific discovery
2022 – Dec 2026 (expected)
B.Tech Metallurgical Engineering & Materials Science + M.Tech Data Science
Indian Institute of Technology Bombay
Ranked 1st in Interdisciplinary Dual Degree Program
2017 – 2022
Honors & Fellowships
2025
Kai Pan & Li Wang Graduate Fellowship
Johns Hopkins University2024–25
Donald S. Rodbell Memorial Graduate Fellowship
Johns Hopkins University2024
Robert B. Pond Sr. Graduate Fellowship
Johns Hopkins University2023
Edwin D. & Rachel Lowthian Endowed Fellowship
Johns Hopkins University2022
Ranked 1st, Interdisciplinary Dual Degree Program
IIT Bombay2017
KVPY Fellow
Dept. of Science & Technology, Govt. of India2015
National Talent Search Scholarship (NTSE)
NCERT, Govt. of India02 · Publications
Selected research
Four peer-reviewed papers spanning materials discovery, perovskite defect physics, and ML frameworks — with more in the pipeline.
Low-energy pathways lead to self-healing defects in CsPbBr₃
First authorDOI
PAL 2.0: A physics-driven Bayesian optimization framework for material discovery
A chemistry-informed machine learning framework for closed-loop material discovery
Deciphering the Dynamical and Structural Factors Controlling Na-Ion Transport in Hydroborate-based Solid Electrolytes
How flyer geometry affects shock wave propagation and spall on impact
First author
03 · Experience
Research & industry
ML Research Intern, CCMS
- Training equivariant graph neural network interatomic potentials (Allegro, MACE, NequIP) on B–H polyhedral anion clusters for large-scale MLIP-MD simulation of solid electrolytes.
- Building an end-to-end ML pipeline — feature engineering, feature selection, and Bayesian optimization — to discover novel hydroborate electrolyte compositions with maximized cation conductivity.
- Applying graph-based sequence modeling to characterize ionic reorientational dynamics over time.
Doctoral Researcher, Clancy Group
- Architected PAL 2.0, a physics-informed active-learning framework combining XGBoost feature selection, Bayesian optimization, and neural-network surrogates — 3× faster convergence than SMAC/Hyperopt across 12+ benchmarks.
- Built hybrid GP + deep-learning surrogate models for high-dimensional search spaces, identifying top-performing perovskite photovoltaics while querying <11% of the design space.
- Engineered a parallelized DuckDB + Python ETL pipeline processing terabyte-scale simulation data (500k+ atoms), automating defect event detection and cutting analysis time by 90%.
- Discovered low-energy self-healing defect pathways in CsPbBr₃ via DFT + NEB — >95% agreement with experiment, published as two first-author papers.
Bachelor's & Master's Thesis
- Trained CNN classifiers to identify NiTi phase transformations from 50+ MD simulations at 90%+ accuracy.
- Modeled Suzuki segregation in FCC Cu with LAMMPS, evaluating stacking fault energies across 50+ impurity configurations.
- Presented findings at ICSMA 19, an international conference in Metz, France.
Computational Science Intern
- Built ML surrogate models on 200+ DFT simulations to predict solid-state battery properties, cutting simulation cost by 40%.
- Benchmarked surrogate predictions against held-out DFT calculations, quantifying uncertainty to prioritize candidates for synthesis.
Computer Vision Intern
- Deployed a YOLOv5-based real-time license-plate recognition system, improving inference throughput by 35%.
- Engineered an automated data preprocessing and augmentation pipeline for variable lighting, occlusion, and camera angles.
Machine Learning Intern
- Built a real-time audio-anomaly-detection pipeline (MFCC/spectral features, SVM/RF/LSTM) at 94%+ accuracy across 10,000+ samples.
- Ran ablation studies across feature counts and architectures to systematically improve the precision-recall trade-off.
Summer Research Intern
- Trained regression and classification ML models on XRD data from 30+ single crystals to estimate impurity concentrations with <5% error.
- Applied PCA and feature engineering over crystallographic structure descriptors to drive feature selection.
Teaching, Mentoring & Service
- Teaching Assistant, 3 graduate courses (Molecular Simulations, Phase Transformations, Computational Modeling) — JHU
- Research Mentor — supervised 4+ undergraduate and 1 graduate researcher
- Lab Manager, Clancy Research Group — coordinated weekly meetings for a team of 10+
- International Student Representative, Materials Graduate Society, JHU
- Mentor, IIT Bombay SMP & DAMP programs — guided 22 students
04 · Projects
Building things
From physics-informed ML frameworks to full-stack AI products — a few things I've built.
DevOps Genie
Multi-agent DevOps automation SaaS with human-in-the-loop infrastructure changes.
- Multi-agent system (LangGraph, GPT-4/Claude) that classifies intent and routes to dedicated debug or infrastructure agents.
- Infra agent generates Terraform/Kubernetes changes validated through custom MCP tool servers, raises a PR, and applies via Atlantis + ArgoCD.
- Shipped full-stack (Next.js 15, Flask, Docker, real-time chain-of-thought UI); grew to 4 customers within 3 months.
05 · Writing
Notes and receipts
Longer writeups behind the charts and analysis I post, with the data and code linked so every number can be checked.
Tech fell 5.5%. Bitcoin fell 1.3%. That is not decoupling.
Week ending Oct 16: the CPI aftermath, bitcoin vs the Mag 7 in dollars, and a fix to Issue 3.
Bitcoin fell 2.3%. Its volume says less than you think.
A quieter week for bitcoin, the Mag 7 comparison, and why trading volume says more about the size of a move than its direction.
Bitcoin's price recovered. Its trading intensity didn't.
Bitcoin made a new price high, but its trading volume has not kept pace with its size or with Nvidia. A look at 200-week dollar trading activity on matched US equity sessions.
Yields up, gold down, Nasdaq at a record
Macro Receipts, issue 1. The 10-year closed the week at 5.28% while gold fell 3.7% and the Nasdaq 100 set a record. Plus bitcoin, the week ahead, and one chart worth the time.
Bitcoin ETF flows: the receipts behind the chart
Every week I post the same bitcoin ETF flow chart. Here is where the numbers come from, how each one gets verified, and what two years of flows actually say.
06 · Contact
Let's talk
Open to research collaborations, ML & scientific computing roles, and interesting problems at the intersection of machine learning and science.