Sandesh Adhikari · Ph.D. researcher, UT Arlington
I build the mathematics and the software that answer that question for systems too large to reason about by hand.
I simulate what could go wrong, teach fast models to stand in for slow ones, and write optimizers that decide where limited money does the most good. I learned it on roads and pipelines facing earthquakes and floods.
Who I am
I studied civil engineering in Nepal, where earthquakes are not hypothetical. Today I work on the question that follows from that: with limited money and an uncertain future, where does protection matter most?
My answer is computational. I build simulations that play out thousands of possible futures, statistical models that learn to imitate slow physics in milliseconds, and optimizers that search through more options than anyone could list. I worked as a software developer before the Ph.D., so I build these tools end to end and I care whether they run fast.
Infrastructure is where I learned this. The shape of the problem is the same wherever capital meets risk: too many choices, incomplete information, and a real cost to being wrong. That is the work I want to keep doing.
Away from the keyboard: hiking, camping, running and the gym. I rarely turn down something I have not tried before.
I build Monte Carlo simulations that carry randomness from the cause all the way to the outcome, and I check that an estimate has settled before I trust it.
Same skill elsewhereScenario generation, risk estimation, valuation by simulation
When one evaluation takes hours, I train a statistical stand-in that answers in milliseconds and reports how sure it is, then spend real computation only where it changes the answer.
Same skill elsewhereCalibrating expensive models, searching strategy and parameter spaces
Thousands of yes-or-no choices, objectives that pull against each other, a budget that never stretches. I use evolutionary search when the problem is a black box and integer programming when it can be solved exactly.
Same skill elsewhereEfficient frontiers, constrained allocation, scheduling
Roads, pipes and supply chains fail as networks. I use graph theory to find the few links everything else depends on, and to shrink a search before it starts.
Same skill elsewhereDependency and contagion structure, supply and transport networks
I shipped production code before I did research: a .NET and PostgreSQL system used to plan road maintenance budgets in Nepal. Now I build research pipelines that chain several simulators together.
LanguagesPython, MATLAB, C, C#, SQL, JavaScript
I have presented research at an ASCE conference, co-authored an open textbook, and been part of a team that delivered hands-on safety training to disaster-recovery workers. A result nobody understands does not get used.
In practiceA conference talk, papers, a book, hands-on training
Track 01 · The problem
This is the decision I work on, shrunk to fit a screen. Every city, utility and fund faces a version of it: more to protect than money to protect it, and a future that only arrives as probabilities.
6 of 6 reinforcements left
Neighborhoods still supplied, on average
Nobody knows which earthquake arrives. A plan has to hold up across thousands of possible ones, so it is judged on a whole range of outcomes and never on one forecast.
The toy above has millions of ways to choose six links. A real network with 300 pipes has more possible plans than there are atoms in the observable universe.
On a real system, scoring one plan means running physics: pressure in every pipe, or traffic on every road. A single run can take minutes to hours, and a search needs thousands.
Track 01 · How I attack it
Play out the uncertainty. Thousands of possible events, each one damaging the system in a different way.
Train a fast statistical model to stand in for the slow simulator, one that also knows where it is guessing.
Search with the fast model in the loop, and spend real simulations only where they could change the decision.
It is the first thing people suggest. For this problem it is the wrong tool on its own, for three reasons.
Machine learning needs many examples. Here the earthquake has not happened yet, so every example has to be simulated, and one simulation can take hours. That buys a few hundred examples, far fewer than a large model needs.
An optimizer hunts for unusual plans, which is exactly where a trained model has seen nothing. Most models answer confidently anyway. I need one that reports its own uncertainty, so I know where the next simulation is worth paying for.
Predicting the damage for one plan does not say which plan to choose out of millions. And a black box cannot be handed to an exact solver, or explained to the people spending public money. A simple model with a known form can do both.
Where plenty of real data exists, machine learning is the right tool. I have used it there, in co-authored work on predicting how long highway construction projects will take.
Track 01 · The work
One first-author paper is in press and listed in full. The others are in peer review, so they appear in outline only. The road-network research is part of the U.S. Department of Transportation project I am working on now.
A systematic review of 183 studies on seismic risk to water and gas pipeline networks, mapping what the field has solved and what it has not: gas systems, secondary hazards such as liquefaction, fire after earthquakes, and decisions made in real time.
Planning repairs for a city road network that has to survive both earthquakes and floods, when the best plan for one is a poor plan for the other.
Deciding where a utility should add shut-off valves, so that a broken main closes a street and not a district.
Recasting a simulation-driven search as a mathematical program that can be solved to proven optimality in a fraction of the time.
Journal of Pipeline Systems Engineering and Practice, ASCE, forthcoming · doi:10.1061/JPSEA2/PSENG-2147
International Conference on Transportation and Development 2026, ASCE, pp. 12–23
Pipelines 2025: Planning and Design, ASCE, Vol. 3, pp. 361–371 · doi:10.1061/9780784486382.039
IISE Annual Conference Proceedings, pp. 1–6
Mavs Open Press, Arlington, Texas · open access
Talks and training
Research gets better when it leaves the building. Two recent occasions when mine did.
I presented our paper on gas pipeline networks and earthquakes. The short version: looped networks generally hold up better after an earthquake than tree-like ones, where one damaged link can cut off everything beyond it.
Read the post on LinkedIn →Three days of hands-on computational training alongside other early-career natural-hazards researchers: high-performance computing, uncertainty quantification, AI and scalable workflows, taught by people from DesignSafe, the NHERI SimCenter, NHERI RAPID and TACC.
Read the post on LinkedIn →Track 02
Before I modeled infrastructure, I tested it by hand: rebound hammers, ultrasonic pulses and ground-penetrating radar on a tower in Kathmandu. This is the engineering that keeps the models honest.
The University of Texas at Arlington
Graduate research assistant at the North Texas Initiative for Resilient Infrastructure, working on the resilience of transportation and pipeline networks. Currently on a U.S. Department of Transportation project on road networks.
Tribhuvan University, Institute of Engineering, Pulchowk Campus
Full four-year merit scholarship. Elective in Time Series Analysis.
Final-year project
Structural analysis and seismic design of a high-rise building.
Time Series Analysis elective
Stochastic generation of river flow sequences from a historical record.
Programming coursework
U.S. Department of Labor Susan Harwood Training Grant · UT Arlington
Part of the team that delivered hands-on, bilingual safety training to resilience workers, and co-author of the twelve-chapter open textbook that grew out of it.
SOFTWEL (P) Ltd., Kathmandu
Building design and surveys for lift and canal irrigation, alongside engineering software on ASP.NET. Contributed to the web system Nepal's Department of Roads uses to track road conditions and plan annual maintenance budgets, and managed its PostgreSQL database.
Research and Training Unit, Pulchowk Campus · for Kathmandu Metropolitan City
Non-destructive evaluation of construction quality: rebound hammer, ultrasonic pulse velocity, rebar location, ground-penetrating radar, ultrasonic tomography and as-built survey, followed by data interpretation.
ADHAAR-2079, Pulchowk Campus
With KTM Geo Lab, in public schools across the Kathmandu Valley
Ecological applications of pervious concrete
Contact
I am always glad to talk about hard problems, new opportunities and research collaborations.
sandeshadhikari34@gmail.com
Sandesh Adhikari