The file format is the easy part: why legacy subsurface data is hard for AI
By Serhii Honcharuk · Published October 27, 2026
Ask a geoscientist where the useful data is and the answer is often “somewhere in the archive”. Decades of well logs, seismic volumes, core reports and drilling records sit in legacy systems, scanned documents and file formats built for the software and storage of their day. AI promises to make all of it searchable in plain language. Before that works, someone has to get it in.
Even the “standard” format is not simple
Digital well logs are the friendliest case. In the mid-1980s the Canadian Well Logging Society set up a committee to replace a patchwork of vendor-specific binary tape formats, many readable only with the originating company’s software. The result, the Log ASCII Standard (LAS), was introduced in 1989. Version 2.0 followed in 1992 to fix inconsistencies, and a more capable version 3.0 arrived around 2000. Today the USGS describes LAS 2.0 as the most widely used digital well-log distribution format, while version 3.0 has seen limited implementation.
Even so, a file that follows the standard still needs interpretation. Curve names differ between service companies and eras. The value that means “no data” varies: common choices include -999.25 and -9999. Units, depth references and header details are filled in by people, with the consistency people bring. Multiply that by thousands of wells and several decades.
The problem is also a funding problem
The same pattern shows up at national scale. The National Geological and Geophysical Data Preservation Program was created in 2005 to help rescue and catalog geoscience collections. At a 2014 congressional hearing, witnesses noted that its authorization of $30 million a year had been matched by only $8 million in total appropriations since 2007. Much of the legacy record is waiting for someone to convert it.
What this means for AI tools
Language models make the front end easy: ask a question, get an answer. They do not remove the need to clean, standardize and verify what sits behind it. Three principles follow:
- Provenance travels with every answer. If a tool cannot say which well, which file and which version an answer came from, an engineer cannot trust it.
- Import is a decision, not just a conversion. Ambiguous units, unfamiliar curve names and missing metadata should go to a person for approval, not be guessed.
- Keep the original. Cleaned data is a derived product. The source file stays untouched and traceable.
The third of our planned product areas is built around this problem: bringing legacy subsurface data into modern platforms with checks at each step, and making it searchable in plain language with sources shown.
Sources
- USGS, “LAS format” (National Geological and Geophysical Data Preservation Program): https://www.usgs.gov/programs/national-geological-and-geophysical-data-preservation-program/las-format
- Kansas Geological Survey, help notes on LAS versions 1.2, 2.0 and 3.0 (including common null values): https://www.kgs.ku.edu/software/SS/HELP/las/index.html
- Handbook history of LAS and earlier binary log formats: https://spec2000.net/08-lasfiles.htm
- Eos (AGU), “Geological data preservation program receives bipartisan support” (2014): https://eos.org/articles/geological-data-preservation-program-receives-bipartisan-support