Why the Loop Matters: Process Variables Are Not Noise
Research Insight — 3D Concrete Printing
Based on data from the RILEM TC 304-ADC International Laboratory Study on 3D-Printed Concrete Mechanics — Bos et al. (2023), open-access dataset. Charts generated by Sunnyday Technologies from the published dataset. See methodology note at bottom for analytical qualifications.
The Conventional Assumption
The conventional approach to 3D concrete printing treats the printing equipment as a neutral delivery mechanism — assume the printer does not matter, test the mix in isolation, report cube compressive strength at 28 days. That assumption is wrong, and a major international study has now quantified exactly how wrong.
The RILEM TC 304-ADC inter-laboratory study on 3D-printed concrete mechanics coordinated testing across multiple independent laboratories worldwide. Each participating laboratory developed its own Portland cement-based mix within the constraints of the study plan — no single mix design was distributed. Labs printed specimens using their own equipment and process parameters and tested under a common protocol. The result is a dataset that captures real-world variation across the 3DCP research community.
What the Data Shows
Across 1,434 specimens from Portland cement-based 3DCP mixes, compressive strength ranged from 3.3 MPa to 130.2 MPa — a 40× spread. The dataset-wide standard deviation was 21.1 MPa. That spread is not primarily measurement error.
Left: Compressive strength distribution by print speed (lab/equipment proxy). Right: The classic water/binder predictor, shown across print speed tiers. Data: RILEM TC 304-ADC ILS-mech (Bos et al., 2023).
What the Variance Actually Comprises
The RILEM ILS dataset is more carefully structured than most multi-lab studies. Alongside the mechanical results, it tracks a range of variables that legitimately explain a significant portion of the observed variance: loading direction (X, Y, Z, and cast reference), curing conditions, extraction location within the printed element, and specimen scale. These are not noise — they are real physical effects that the study was designed to quantify.
Orientation alone is a substantial predictor: specimens tested in the Z (build) direction consistently underperform those tested in X or Y due to interlayer bond geometry. Curing regime affects hydration state at test. Scale effects are well-established in concrete testing. When these variables are controlled or accounted for, within-condition variance is meaningfully reduced.
But here is the key finding: even after accounting for direction, curing, extraction location, and scale, substantial inter-laboratory residual variance remains. That residual is not explained by the tracked experimental variables. It is driven by equipment geometry, print speed, layer height, open time, and the interaction between process parameters and the mix — variables that differ between labs and are not fully captured in the dataset.
The Left Chart: Equipment Identity as a Signal
The left panel groups specimens by print speed — each speed cluster is a proxy for a different laboratory and equipment configuration in the ILS. The spread within and between groups is substantial. Labs running at 80 mm/s produce a meaningfully different population of results than labs running at 200 mm/s, even accounting for the fact that these labs also developed different mixes.
The variance is real, it is systematic, and a meaningful portion of it is traceable to process parameters that vary between labs but go unrecorded or uncontrolled in aggregated literature datasets.
The Right Chart: The Classic Predictor in Context
The water/binder ratio has been the dominant predictor of concrete compressive strength for over a century. Abrams’ Law works reliably for cast concrete because the production process is standardized — vibration, formwork, curing conditions. In additive manufacturing, process is not standardized. It varies between every lab, every printer, and every print session.
The right panel plots w/b ratio against compressive strength, coloring specimens by print speed tier. The classic inverse relationship is visible within tiers — but the location of that relationship shifts substantially between speed groups. The trendlines diverge in both slope and intercept.
The correlation between w/b ratio and compressive strength across the full pooled RILEM dataset is R ≈ 0.35. This is computed across all loading directions, curing conditions, specimen scales, and extraction locations combined — the most conservative possible estimate. Water/binder ratio explains roughly 12% of total pooled variance. The remaining 88% is shared between known test variables (direction, curing, scale, location) and untracked process variables. The known variables are large; the untracked process residual is meaningful and, critically, systematic.
The Implication for Machine Learning
This finding has a direct and underappreciated implication for any ML model trained on 3DCP data. The distinction matters: tracked variables are features; untracked variables are confounders.
Loading direction, curing regime, specimen geometry, and extraction location are legitimate model inputs. A well-structured 3DCP dataset that captures these variables gives a model real signal to learn from. The RILEM dataset demonstrates this clearly — these are variables worth encoding.
The problem is different: it is the equipment and process variables that are not tracked — print speed, layer height, nozzle geometry, open time management, pumping configuration — that become hidden confounders. When a model is trained on aggregated results from many labs without these variables recorded, it cannot distinguish material response from equipment response. It learns a relationship that is partly composition, partly process artifact, with no way to separate them at inference time.
The solution is not to avoid multi-lab data. The solution is to make process variables known inputs rather than hidden confounders. The RILEM dataset’s tracked variables — direction, curing, scale — demonstrate exactly this: when you record it, you can use it. The closed-loop Sunnyday system extends this principle to the process variables that literature datasets leave unrecorded.
What This Means for Sunnyday Technologies
The RILEM data makes a precise case: formulation alone explains ~12% of variance when all conditions are pooled. A large portion of the remaining variance is attributable to known, trackable test variables — direction, curing, scale, location. A meaningful residual is driven by equipment and process parameters that literature datasets do not capture. That residual is the signal the Sunnyday closed loop is specifically designed to isolate.
CEMFORGE generates formulations. M3-CRETE prints them under repeatable, documented process conditions — with print speed, layer height, open time, and equipment geometry recorded as part of every test cycle. Physical specimens are tested. The results feed back into CEMFORGE with process parameters attached, not averaged away.
One completed predict–print–test cycle on a single controlled system provides training signal that literature aggregation cannot replicate — not because of scale, but because of information completeness. The process variable that the RILEM study identifies as a major source of unaccounted variance becomes, in the Sunnyday loop, a known model input.
Chart methodology: 1,434 3D-printed specimens filtered to 28-day tests (±3 days). Each participating laboratory developed its own Portland cement-based mix per study constraints; mixes were not identical across labs. Print speed used as lab/equipment identity proxy; each speed cluster corresponds to a distinct laboratory configuration. Speed tiers defined by tertile split of the speed distribution. Trendlines are ordinary least-squares linear fits per tier. R≈0.35 is computed across all pooled conditions (directions, curing regimes, scales, extraction locations) and represents a lower bound on the explanatory power of w/b ratio under controlled conditions.