# Dixon Glacier Model Workbench

A glacier model, opened up so the decision can be checked. Research tooling · public. 2026. Role: Model author and sole developer.

Published 2026-07-23 · Updated 2026-09-30 · Author: Kai Myers, K.AI Consulting, Anchorage, Alaska
Canonical page: https://kaiconsulting.ai/work/dixon-glacier

## In one sentence

The Alaska Energy Authority is pursuing a diversion that depends on Dixon Glacier's future meltwater. I built the model that projects it, then built a workbench that shows its data, its assumptions, and its uncertainty rather than just its answer.

## The problem

The Alaska Energy Authority is actively pursuing the Dixon Diversion: a dam near the glacier's toe and a tunnel of roughly 4.9 miles carrying meltwater to Bradley Lake, the state's largest hydro plant. It would lift Bradley's annual output by about 50 percent (roughly 60 MW, power for up to 30,000 more homes) at a cost of hundreds of millions of dollars.

The case for it rests on how much water Dixon will produce and for how long. A glacier's runoff rises as it melts faster, then falls permanently once there is not enough ice left to sustain the increase. That turning point is peak water, and whether it is decades away or already behind us is central to whether the diversion pencils out.

A number alone does not settle that. What a decision of this size needs is the number plus everything behind it: which observations constrained it, which choices the modeler made, and how wide the uncertainty really is.

## What was built

A bespoke daily glacier mass-balance model, written from first principles rather than adapted from an existing package, calibrated with Bayesian MCMC against stake measurements, geodetic mass change, and observed snowlines.

A public workbench built around eleven stations, each one a different instrument on the model: a geospatial explorer, an atlas of every input dataset with its provenance, the modeling decisions and why each was made, how the model works, calibration, validation, a projections explorer, and reproducibility.

An AI disclosure station documenting how the model was built with LLM assistance and where that assistance was and was not trusted.

A retrieval-backed assistant that answers questions against the model's own corpus, so a reader can interrogate the thing directly instead of taking the summary on faith.

## Outcome

- Median peak water of 2040, 2061, and 2083 under SSP1-2.6, SSP2-4.5, and SSP5-8.5
- 99.75% of model runs put peak water after 2025, so the turning point is still ahead
- Every input dataset published with source, units, CRS, coverage, and known caveats
- Live at dixonglacier.com as a public, checkable companion to the thesis

## Stack

Python, emcee (Bayesian MCMC), React, Vite, TanStack Router, bun, Cloudflare, Claude API

Live: https://dixonglacier.com