Autoencoders for Seismic Waveforms: Compress, Learn, and Generate
Autoencoders learn compact representations by reconstructing seismic data. This tutorial explains their use...
Utpal Kumar, PhD
Earthquakes right now
Magnitude 2.5+ over the past 24 hours.
Loading recent earthquakes…
- 150+ Posts
- 12 Topics
- Oct 2026 Updated
RLHF Explained: How Human Feedback Shapes AI Agents
How human comparisons become a training signal, where DPO fits, and why useful AI agents still need indepen...
Recent posts
Recent technical articles on seismic analytics, scientific computing, numerical modeling, and production data workflows.
Finding tiny earthquakes by their fingerprints
A three-station seismic matched-filter experiment shows how repeating waveform shapes reveal small earthquake candidates, and why strong matches still need v...
What ai agents can learn from the bayesian brain
How neuroscience, autonomous laboratories, and language-model research can inform agents that preserve evidence and revise their beliefs.
From github to zenodo: make your code and data citable
Turn a GitHub release into a citable research archive, share supporting data, and choose the right DOI for your paper.
How ai agents choose what to learn next: expected information gain explained
How expected information gain helps AI agents choose the next observation, with Shannon entropy, a worked building-monitoring example, and practical limits.
Agentic harness explained: how ai agents get work done
An agentic harness gives an AI model tools, context, memory, permissions, and feedback. See how the pieces work together in a scientific Python workflow.
When frequency changes signal structural damage
An unexpected drop in a building’s natural frequency is the structure reporting that it has lost stiffness — usually the first sign of damage. Here is why to...
From phase pickers to seismic foundation models
Foundation models ask whether one pretrained representation of seismic waveforms can support picking, detection, classification, and other downstream tasks—w...
Caffeinate on and off controls for macos with alfred and swiftbar
A compact macOS setup with two Alfred workflows for turning caffeinate on and off, plus SwiftBar for keeping the current status visible in the menu bar.
Automatic p/s phase picking with deep learning using seisbench
Load a waveform with ObsPy, run a pre-trained PhaseNet (or EQTransformer) model from SeisBench, and plot the picked P and S arrivals — a hands-on bridge from...
Markov chain monte carlo for geoscientists
An intuitive guide to MCMC: a walker in the fog whose footprints map the posterior. We build a Metropolis sampler from scratch, check it against an exact ana...
Probabilistic earthquake location with nonlinloc
How NonLinLoc locates earthquakes by mapping a full probability density over the search volume instead of iterating to a single point — the Bayesian idea, th...
Modern seismic monitoring systems: how ml and the cloud build earthquake catalogs
A tour of how modern earthquake-monitoring systems turn continuous seismic waveforms into earthquake catalogs — the ML pipeline of picking, association, and ...
Finding tiny earthquakes by their fingerprints
A three-station seismic matched-filter experiment shows how repeating waveform shapes reveal small earthquake candidates, and why strong matches still need v...
What ai agents can learn from the bayesian brain
How neuroscience, autonomous laboratories, and language-model research can inform agents that preserve evidence and revise their beliefs.
From github to zenodo: make your code and data citable
Turn a GitHub release into a citable research archive, share supporting data, and choose the right DOI for your paper.
How ai agents choose what to learn next: expected information gain explained
How expected information gain helps AI agents choose the next observation, with Shannon entropy, a worked building-monitoring example, and practical limits.
Agentic harness explained: how ai agents get work done
An agentic harness gives an AI model tools, context, memory, permissions, and feedback. See how the pieces work together in a scientific Python workflow.
When frequency changes signal structural damage
An unexpected drop in a building’s natural frequency is the structure reporting that it has lost stiffness — usually the first sign of damage. Here is why to...
From phase pickers to seismic foundation models
Foundation models ask whether one pretrained representation of seismic waveforms can support picking, detection, classification, and other downstream tasks—w...
Caffeinate on and off controls for macos with alfred and swiftbar
A compact macOS setup with two Alfred workflows for turning caffeinate on and off, plus SwiftBar for keeping the current status visible in the menu bar.
Automatic p/s phase picking with deep learning using seisbench
Load a waveform with ObsPy, run a pre-trained PhaseNet (or EQTransformer) model from SeisBench, and plot the picked P and S arrivals — a hands-on bridge from...
Markov chain monte carlo for geoscientists
An intuitive guide to MCMC: a walker in the fog whose footprints map the posterior. We build a Metropolis sampler from scratch, check it against an exact ana...
Probabilistic earthquake location with nonlinloc
How NonLinLoc locates earthquakes by mapping a full probability density over the search volume instead of iterating to a single point — the Bayesian idea, th...
Modern seismic monitoring systems: how ml and the cloud build earthquake catalogs
A tour of how modern earthquake-monitoring systems turn continuous seismic waveforms into earthquake catalogs — the ML pipeline of picking, association, and ...