A research team led by Jaewook Jeong, an associate professor at Seoul National University of Science and Technology, has published an AI pipeline that reads a building’s structural calculation package and its structural drawings, matches every beam and column across the two, and flags where they disagree. The paper, by Minsang Gu, Jaewook Jeong and Louis Kumi, went online in Engineering Applications of Artificial Intelligence on October 13, 2026, and the university announced it on November 3. On a 19-story reinforced concrete apartment building in Seoul, the system found 253 discrepancies, and the authors class 29 of them as critical to structural stability. The workflow it targets is structural document review, the check of rebar sizes, bar counts, stirrup spacing and member dimensions between what the engineer calculated and what the drawings tell the field to build.
What it does
Two documents go in. The calculation package comes out of analysis and design software (MIDAS ADS and MIDAS SDS on the case project) and lists each member’s loads, reinforcement and layout. The structural drawings are the schedule-type sheets a drafter builds from that package in CAD. The paper’s premise is that this hand transfer between two formats, often by different people, is where omissions and typos get in.
The pipeline then runs in three stages:
- It collects the documents and drops low-quality pages.
- It pulls member data out of each document. On the calculation pages, a layout model called DocLayNet-YOLO finds the regions that hold member information, and Surya-OCR, an optical character recognition model that turns a page image into text, reads them. On the drawings, the software finds the member codes first, then uses OCR with edge and line detection to locate the schedule tables, whose rows and columns vary in width.
- It matches each member across the two documents by member code and direction, compares reinforcement and dimensions, and sorts every disagreement by type.
Results come out in a web application the team built. The method works from the documents alone, with no BIM model, and the authors aim it at small and mid-size projects that do not build one. They name structural engineers, design consultants, contractors and construction supervision teams as the people who would use it. The paper also notes that US practice splits structural documents among architects, structural engineers and CAD technicians, which is the kind of divided authorship the check is built for.
Evidence so far
The evidence is one case study the authors ran on the Seoul project, reported in the paper’s abstract and the preview ScienceDirect shows of the full text. The authors declare no competing interests, and a Korean government grant for commercializing research outcomes funded the work.
Extraction accuracy, the share of fields the software read correctly, was 95.56% on the calculation package and 90.47% on the drawings. The 253 discrepancies were all in beams and columns.
| Discrepancy type | Count |
|---|---|
| Missing entries | 224 |
| Main reinforcement mismatches | 15 |
| Dimension mismatches | 8 |
| Stirrup mismatches | 3 |
| Duplicate entries | 3 |
| Total | 253 |
So the 29 critical errors are the 253 minus the 224 missing entries. Every mismatch in bar size, bar count, stirrup spacing or member height, plus the three duplicates, landed in the critical group.
| Critical error | Count |
|---|---|
| Dimension height | 8 |
| Bottom rebar diameter | 6 |
| Top rebar diameter | 3 |
| Top rebar count | 3 |
| Bottom rebar count | 3 |
| Stirrup spacing | 3 |
| Duplication | 3 |
| Total | 29 |
The two accuracy figures matter for reading the larger number. If the software misread about one drawing field in ten, some of the 224 missing entries could be fields it failed to read and not gaps in the drawings, and the preview does not show how the authors confirmed each flag. The time the check took, and a manual review of the same documents for comparison, are not in the abstract or the preview. (The paper cites conflicting design documents as a contributing factor in the New World Hotel collapse in Singapore and an underground parking structure collapse in Incheon.)
How to measure it on your projects
The metric to borrow is confirmed discrepancies per structural sheet, counted before rebar detailing starts and sorted into the paper’s five types. Count a flag only after the engineer of record or your detailer confirms it, so OCR misreads do not inflate the total. Dividing confirmed flags by all flags gives you the tool’s precision, the share of its flags that turned out to be real errors.
You need the calculation package for this, and the engineer of record does not always issue it to the contractor. Ask for it at the start of the pilot.
Capture two baselines on a recent concrete job: the RFIs that traced back to a conflict between the calculations and the beam or column schedules, and the hours your detailer or project engineer spent checking schedules by hand. A first step on one job is to take a single floor’s beam schedule and the matching calculation pages, check them by hand, and log the hours and each discrepancy by type. Run the tool on the same floor afterward and compare both counts and both times.
Availability and cost
This is a research system. The paper describes a web application, and the abstract and preview do not mention a public code release, a commercial version or a price. The full text sits behind Elsevier’s paywall.
Earlier posts covered commercial drawing review tools, including Buildcheck, Document Crunch’s drawing review and Bluebeam Max.