The Loop is the Product
3D concrete printing has been treated as three separate trades — mix design, machine building, and physical testing — that hand off to each other through paper specs and human translation. We treat it as a single closed development loop, with each stage feeding the next as structured data.
Three stages, one feedback loop.
Most concrete-printing development today looks like this: a mix is designed in isolation, handed to a printing crew, run through a machine whose parameters are not fully documented, tested as cured cubes that don’t reflect print anisotropy, and reported in a paper that doesn’t feed back into the formulation.
That sequence is open-loop. Every iteration starts from approximately the same place. Improvements come from individual labs making local progress, but the field-wide signal is lost in the handoffs.
We design the system as a closed loop instead:
- Predict — CEMFORGE generates candidate mix designs from your available materials and target properties.
- Print — M3-CRETE prints the specimens with documented, repeatable process parameters.
- Prove — physical testing validates the prediction at the orientations and rates that match the print direction, capturing anisotropy.
The loop closes when the validated test results feed back into CEMFORGE’s training set, calibrated to the specific materials, printer, and operating conditions that produced them. Each cycle tightens the model’s accuracy for that operator.
Process variables are not noise.
The RILEM TC 304 inter-laboratory study on 3D-printed concrete mechanics ran the same mix design through different printers at different labs and measured the spread. The mechanical properties varied by an order of magnitude — not because of measurement error, but because layer height, print speed, open time, and nozzle geometry are first-class drivers of outcome variance, not background noise.
What that means for AI formulation: if your training data mixes results from many labs, each with different process parameters that were never recorded, you are training on process noise as if it were materials signal. The model learns something that is partly real and partly an artifact of which machine produced which specimen.
The printer is not a neutral delivery mechanism. It is a primary variable, and structured training data has to capture it.
Sunnyday operates the formulation engine, the printer, and the test program. The data each specimen produces is structured to retain the process parameters that generated it. That coupling — software plus hardware plus testing, designed as one system — is the data asset that no single component on its own can produce.
The market this approach was built for.
The dominant approach in industrial concrete is appropriate for industrial concrete: large players, large mixes, large fixed-asset projects, with established testing and qualification infrastructure. That path is real, valid, and well-funded.
It does not reach everyone. Small printers, regional contractors, university labs, cement-plant new-product groups, makers, researchers, and humanitarian programs operate at a scale where commercial mix qualification is economically out of reach and proprietary printers are opaque to the people who’d most benefit from inspecting them.
Our approach is built for that under-served set. Open hardware on the print side, AI formulation that runs on a laptop on the predict side, structured testing that gets cheaper per specimen as the dataset grows on the prove side. The economics of getting to a validated, printable mix change when the development cycle is something any small operator can run.
Where this work sits in the literature.
The methodological lineage is Olson’s materials-by-design framework (Olson, 1997, Science 277:5330:1237; Xiong & Olson, 2015, MRS Bulletin 40:5:1035) and the broader Integrated Computational Materials Engineering (ICME) program. The Modified Andreasen-Andersen particle-packing model (Funk & Dinger, 1994) is the literal backbone of CEMFORGE’s gradation theory.
Sunnyday’s contribution is the closed-loop digital twin: chemistry plus particle-size distribution plus process parameters plus environment, anchored to the open Open3DCP printability schema, with structured feedback from physical testing. ML now, ANN trajectory.
For the full methodology, see our research at /research/ and the open Open3DCP schema at open3dcp.org.
Want the full methodology paper?
Convergence Engineering, the long-form methodology paper behind this approach, is open-access on the research hub.