ShellBench

How much do thin-shell floors cut the whole-building A1–A3 embodied carbon of a GB 50010 / GB 50011 RC frame across 3–8 storeys, and how does the saving change with storey count and seismic intensity once diaphragm, tie and topping needs are met?

(Revised 22 Oct 2026; see docs/research_question.md.)

A thinner floor weighs less. Less weight means smaller seismic forces, which can mean smaller beams, columns and footings. ShellBench follows that chain through the whole building, using real design codes, rather than comparing floors on their own.

Status (30 Oct 2026, Day 22). Week 1 is complete (review). The anchor model is built and checked; the shell floor work has started. Daily progress is in LOG.md. First release: v0.0.1, "Baseline spec and gravity check".

Area	State
Thesis audit and validation targets	Done. 16 thesis inconsistencies (B1–B16) each have a recorded decision.
Anchor frame (OpenSeesPy)	Dead, live and combined gravity load, wind, the GB 50011 base-shear method and a response-spectrum check with a 3D model are in src/shellbench/. Every difference above 10 % from the thesis has a written reason (docs/anchor_notes.md).
Design checks and quantities	GB 50010 / GB 50011 checks of the beams and columns, a quantity take-off and an A1–A3 carbon step are in the package.
Shell floor	Concept approved (layout A, Day 22; docs/floor_concept.md). A CalculiX smoke test differs from the closed form by 0.005 %.
Tests	pytest: 132 passed (31 Oct 2026).
Open	The GB 50011 / JGJ 3 topping rule is still [?] and is needed before Day 27.

Plan: the build ends 13 Nov, the author checks 14–17 Nov, v1.0 follows on 17 Nov, and Paper 1 is submitted by 22 Nov.

---

Approach

The study is checked against one real, fully documented building: the author's own thesis design. That building is validation case 0. Every model must reproduce it before it is trusted on anything else. The result is the generator, the building family and the floor interface, which anyone can run on their own building or floor (see docs/impact_strategy.md).

flowchart LR
    T["Thesis building<br/>(4-storey RC frame)"] --> A["Audit<br/>baseline spec"]
    A --> V["Validation<br/>targets"]
    V --> M["Anchor model<br/>OpenSeesPy"]
    M --> G["Generator<br/>building family"]
    G --> F["Design each building twice<br/>flat slab vs shell floor"]
    S["Shell floor module<br/>form-finding + FE check"] --> F
    F --> P1(["Paper 1<br/>whole-building carbon"])
    F --> D[("Open dataset")]
    D --> AI["Surrogate model"] --> P2(["Paper 2<br/>does AI pick the same floor?"])

    classDef done fill:#c8e6c9,stroke:#2e7d32,color:#000;
    classDef doing fill:#fff3c4,stroke:#b8860b,color:#000;
    classDef todo fill:#eeeeee,stroke:#888,color:#000;
    class T,A,V,M done;
    class S doing;
    class G,F,D,AI,P1,P2 todo;

Loading

Green: done. Yellow: in progress. Grey: not started.

1. Anchor. Rebuild the thesis building in code (gravity, wind, earthquake), add the code checks and the embodied-carbon count, and match the thesis results.
2. Generator. Write code that designs a building the way the thesis did, prove it on the thesis sizes, then use it to produce a family of buildings.
3. Shell floor. Form-find a vaulted floor panel, check it with a benchmarked shell finite-element solver, and size it for thickness, ties and buckling.
4. Comparison. Design every building in the family twice, once with a flat slab and once with a shell floor, and explain the saving term by term.
5. Surrogate. Train a fast model on the family and test whether it picks the same floor as the full design.

Roadmap

Each week has a GitHub issue with its goals.

Week	Dates (2026)	Goal	Release	Status
1	19–25 Oct	Audit the thesis and model it under gravity	v0.0.1	Done
2	26 Oct–1 Nov	Finish the anchor: lateral loads, code checks, carbon	v0.1 (31 Oct)	In progress
3	1–7 Nov	The generator and the building family	v0.2	Not started
4	30 Oct–6 Nov (alongside week 3)	The thin-shell floor module	v0.3 (5 Nov)	In progress (concept approved)
5	7–13 Nov	Whole-building results	dataset DOI (11 Nov)	Not started
6	4–22 Nov	Write and submit Paper 1	submitted by 22 Nov	Not started
7	7–13 Nov	The AI surrogate study	v0.5	Not started
8	17 Nov–13 Dec	Paper 2 and the public package	v1.0 (17 Nov)	Not started

Dates moved on 28 Oct 2026 with the author's approval: the build ends 13 Nov, 14-17 Nov is the author's check window, and no check is dropped. The day-by-day map is docs/schedule.md.

The hypotheses (H1–H4) will be frozen in public in Week 3, before any family result exists.

Reuse items built into the weeks above (details in docs/impact_strategy.md): carbon factors as data files and a Zenodo DOI for every release from v0.1 (Week 2); buildings defined as data, not code (Week 3); a written floor plug-in contract with a conformance test (Week 4); reference problems and a datasheet in the dataset (Week 5); one-command reproduction of each paper's figures (Weeks 6 and 8); JOSS-ready documentation and two tutorials in v1.0 (Week 8).

Done so far

Item	What it holds
docs/baseline_spec.md	Every model input for the thesis building, each with its thesis section or table. 16 inconsistencies found in the thesis (B1–B16), each with the author's decision and reason.
validation/thesis_values/	Eight CSV files of thesis results: beam-end moments, column axial forces, D-values, wind drift, seismic storey shear and drift, beam steel, the full live-load distribution table and the roof rows of the combination table. Every row names its source table. Flags D1–D21 mark values that look wrong; they are left as printed.
tests/	Checks that every validation file has units in its headers, no empty cells and numbers where numbers belong; closed-form and equilibrium tests of the anchor frame.
src/shellbench/anchor_frame.py	The thesis transverse frame as a 2D elastic OpenSeesPy model, loaded with dead and live load the way the thesis distributes them.
src/shellbench/combinations.py	The thesis load combinations as data (1.3D + 1.5L now; wind and seismic ones wait for their load cases).
src/shellbench/thesis_checks.py	Arithmetic checks on the thesis tables: column sums and joint balance of the live-load moment distribution, factors of the combination table.
results/	Dead-load, live-load and 1.3D + 1.5L comparisons with the thesis (beam-end moments, column axial forces, % differences, check runs), and the thesis-table checks. Causes of every difference above 10 % in docs/anchor_notes.md.
docs/gap_matrix.md	23 papers from 2020-2026, each with what it studied, whether seismic loads and ties were included, and a section locator; gap statements G1-G4. Sources in docs/reading_list.md, all legal open copies.
docs/goal_stress_test.md	Red-team review of the research question (holes H1-H9) and the fixes adopted.
docs/week01_review.md	Week 1 review: every Done-when line checked against a file, and the weak points to handle before the anchor report.
notes/gb50011_base_shear_primer.md	Study note on the GB 50011 base-shear method worked on the thesis building, with the clauses to check before Day 9.
docs/impact_strategy.md	How others will reuse and cite ShellBench: what the charter already covers, additions A1-A10 placed in Weeks 2-8 without moving a date, and ideas parked under Later.
docs/floor_concept.md	The approved shell floor concept (layout A), levelling options and the f/L range 1/8 to 1/20.
src/shellbench/	Also holds the wind, seismic, response-spectrum, 3D anchor, design-check, carbon and CalculiX modules, each with tests.
docs/q1_standards.md	Journal and publication-ethics rules (Elsevier, Springer Nature, COPE) that the project follows every day, with a daily compliance check.

The anchor building

A four-storey cast-in-place RC moment-frame office building in Panzhihua, Sichuan, China, designed by the author as a 2023 thesis. Values below are from docs/baseline_spec.md; the ID in brackets is the row there.

Property	Value
Transverse spans	7.2 m + 6.6 m + 7.2 m (G3)
Storey height	3.6 m typical, 4.55 m ground storey (G6, G7)
Columns	600 × 600 mm (M5)
Frame beams	300 × 600 mm side spans, 300 × 500 mm corridor (M1, M2)
Slab	120 mm (G10)
Materials	C40 concrete, HRB400 bars (X1, X5)
Seismic	Intensity 7 (0.10 g), group 1, site class II, Tg = 0.35 s, seismic grade three (S1–S5)
Design codes	Chinese codes GB 50009 (loads), GB 50010 (concrete), GB 50011 (seismic)

The thesis is kept private and is not in this repository. Only the values typed from it are.

Repository map

ShellBench/
├── src/shellbench/       Python package (anchor frame, wind, seismic, carbon, CalculiX)
├── tests/                pytest suite
├── docs/                 specifications, literature, charter and journal-rule notes
├── validation/
│   └── thesis_values/    thesis results as CSV validation targets, with flags
├── data/                 input data (carbon factors)
├── results/              generated results (anchor frame, gravity)
├── figures/              generated figures
├── paper/                manuscripts (empty so far)
├── notes/                working notes (GB 50010 and GB 50011 primers)
├── .claude/commands/     /sync and /today commands
├── LOG.md                daily research log
├── CLAUDE.md             rules for AI assistance
├── environment.yml       conda environment "research"
├── pyproject.toml        package and pytest configuration
├── CITATION.cff          citation metadata
└── .zenodo.json          Zenodo release metadata

Getting started

Requires conda. The environment pins Python 3.12, which OpenSeesPy needs.

git clone https://github.com/syednafizreja/ShellBench.git
cd ShellBench
conda env create -f environment.yml
conda activate research
pip install -e .
pytest

The environment includes OpenSeesPy (frame analysis), gmsh (meshing) and CalculiX (shell finite elements), plus the usual numpy, scipy, pandas and matplotlib.

Working rules

• SI units everywhere. Every CSV header states its unit, e.g. M_end [kN*m].
• Every number traces to a source: a thesis section or table, a code clause, or the code in this repository.
• Thesis errors are flagged, not fixed. Values are typed as printed, and disputed ones are settled by a recorded decision.
• Judgement is marked. Anything that is reasoning rather than a sourced value carries [J].

Use of AI assistance

Code and documents in this repository are drafted with Claude (Anthropic) and checked by the author. All engineering decisions are the author's. The rules Claude works under are in CLAUDE.md.

How to cite

Citation metadata is in CITATION.cff; GitHub's "Cite this repository" button uses it. The repository is public, and Zenodo mints a DOI for each GitHub release, so only final files are released.

Author

Syed Nafiz Reja Hasib

Licence

MIT. See LICENSE.