3D Concrete Printing Mix Design: From Particle Packing to Print Window, A Complete Engineering Guide
3D Concrete Printing Mix Design: From Particle Packing to Print Window, A Complete Engineering Guide
Extrusion 3D concrete printing (3DCP) has matured from laboratory curiosity to deployed construction technology in roughly fifteen years. Gantry printers from a dozen manufacturers now routinely produce walls, retaining structures, and habitable shells. What has not matured at the same pace is mix design. The fresh mortar that flows through a hose, exits a nozzle, supports its own self-weight, bonds to the layer below, and develops adequate hardened strength is still the binding constraint on what 3DCP can do, where it can do it, and at what embodied carbon. This article is a working engineer’s review of that problem. We assume the reader knows what a Bingham fluid is, has seen a slump cone, and can read a hydration curve. We do not assume the reader has worked in 3DCP specifically. Where helpful we point to the standards work emerging from RILEM Technical Committee 304-ADC, among other available reference frameworks, and to specific peer-reviewed studies that anchor the engineering claims we make. We treat these as points of compatibility, not commitments; newer measurement protocols enter our analysis on equal footing as they mature.
1. Why mix design is the binding constraint in 3DCP
The hardware problem in 3DCP is essentially solved. Off-the-shelf gantries and articulated arms can position a print head with submillimeter repeatability over volumes large enough to print a single-family house in a single setup. Pumps with adequate pressure margin, progressive cavity or piston, are commercial. Inline accelerator dosing, real-time path planning, and closed-loop layer height control are all in the field. By contrast, the formulation that sits inside the hose is bespoke, brittle to material substitution, and routinely the cause of print failures, mechanical underperformance, and project delay (Roussel, 2018; Buswell et al., 2018).
The reason is structural. A printable mortar must satisfy several rheological constraints simultaneously, all of which evolve in time after water contact, and most of which are coupled. It must remain pumpable through a hose for the duration of a working batch. It must extrude cleanly through a nozzle at imposed shear without phase separation or filter pressing. It must develop yield stress fast enough to support its own self-weight at the layer-cycle time the printer imposes. It must retain enough surface fluidity to bond to the layer above and below. And it must not slump, deform, or crack as drying and hydration shrinkage compete with stiffness gain. A mix that wins on any one of these axes can lose on another. The “printability window” is the narrow corridor in time and parameter space inside which all conditions hold simultaneously (Wangler et al., 2016).
In practice, most field-reported 3DCP failures, cold joints, layer collapse, voiding at the interlayer, dimensional drift, are mix-state failures rather than equipment failures. RILEM TC 304-ADC has begun to quantify this through coordinated interlaboratory work (one of several available reference points), and a recent database system spanning multiple participating labs documents that even when the same nominal formulation and printer settings are used, between-lab variance in hardened mechanical properties remains the dominant source of scatter (Robens-Radermacher et al., 2025). The implication is that mix design is not a parameter to be tuned at the end of a project but the project. Everything downstream, structural sizing, surface finish, layer cycle, embodied carbon, depends on a formulation that lands and stays inside the printability window for a sample of local materials that the production crew can actually source.
There is a second-order reason mix design is the binding constraint: 3DCP discards two of the strongest correction tools available to conventional concrete, vibration and formwork. A poorly placed conventional concrete can be vibrated, troweled, and screeded; a poorly extruded 3DCP filament cannot be retrieved. There is no formwork to enforce geometry, and no vibrator to close voids at the interlayer. Every defect that appears in the printed wall is a defect that the mortar created on its way out of the nozzle. The mix has to do, fresh, what formwork and vibration do for cast concrete. That shifts the entire quality-control burden upstream into rheology and gradation, which is where the rest of this article focuses.
2. ICME applied to 3D concrete printing
The intellectual lineage for what CEMFORGE does sits in materials science, not in software engineering. Gregory B. Olson, often called the father of materials design, formalized in the late 1990s a closed-loop systems approach in which composition, processing, structure, and performance are linked computationally and traversed in both directions: properties drive structure, structure drives processing, and processing drives composition (Olson, 1997; Olson, 1998). At QuesTek Innovations, Olson’s group operationalized that loop into a proprietary platform branded Materials by Design and used it to qualify alloys for aerospace and defense, anchoring what is now the discipline of Integrated Computational Materials Engineering (ICME). The methodology has since been picked up across alloy design, ceramics, and increasingly cementitious systems.
3DCP mix design is the canonical ICME problem expressed in cement chemistry. Composition is the binder blend, the SCM, the filler, the gradation, and the admixture chemistry. Processing is the mixer, the pump, the nozzle, the layer-cycle time, and the ambient conditions. Structure is the microstructure that emerges from hydration, packing, and shear history. Performance is fresh-state printability, hardened mechanical strength, durability, and embodied carbon. None of these can be optimized independently. Holding processing constant and tuning composition will improve one performance axis and silently degrade another. ICME’s contribution, and Olson’s, is to insist that the loop is closed and that the design tool walks it explicitly rather than treating it as four separate problems.
CEMFORGE inherits this stance directly. Its Predict, Print, Prove loop is the ICME loop renamed for the 3DCP context: predict candidate mixes against composition and processing constraints, print verification specimens that close the loop on structure and fresh-state behavior, and prove the formulation on the hardened performance and embodied-carbon axes. The hand-off between predict and prove is captured in the Open3DCP open data schema, which keeps the four ICME nodes in a single record per specimen so the loop can actually iterate on data rather than on memory.
3. The five regimes of 3DCP rheology
RILEM TC 304-ADC and its predecessors have been working since the late 2010s on a shared vocabulary for fresh-state 3DCP behavior, and the broader literature is converging similarly. The literature distinguishes at least five property regimes, each with a characteristic rheological metric (Buswell et al., 2018; De Schutter et al., 2018; Mechtcherine et al., 2020). The vocabulary below is what the field uses today; we treat it as compatible reference framing rather than a fixed methodological commitment, and remain open to refinements as newer measurement methods mature.
Pumpability is the ability of the fresh mortar to be transported from the mixer to the print head under imposed pressure without phase separation, blockage, or filter pressing. The relevant metric is plastic viscosity at the shear rate experienced inside the hose. The failure mode that ends most pumping trials is filter pressing, the loss of free water into the porous mortar plug ahead of the piston, which is governed by the permeability of the granular skeleton and by the dynamic yield stress of the suspension (Secrieru et al., 2020; Mechtcherine et al., 2020).
Extrudability is the property of a mortar to be forced through the nozzle aperture as a continuous, dimensionally stable filament. The relevant metric is the dynamic yield stress measured under controlled shear. Above some chemistry-dependent ceiling, extrusion produces ragged or torn filaments unless pressure is increased, which then risks pump or hose limits (Tay et al., 2019; Roussel, 2018).
Buildability is the ability of a deposited filament to support the self-weight of subsequent layers without buckling, slumping, or deforming. The metric is the static yield stress and its rate of growth with rest time, the structural buildup rate (with units of pascals per second). Mixtures with insufficient structural buildup rarely build more than a few courses at typical layer-cycle times. Mixtures with excessive structural buildup tend to lose layer adhesion (see below) (Roussel, 2018; Perrot et al., 2016; Reiter et al., 2018).
Layer adhesion is the tensile bond developed at the interlayer interface between sequentially deposited filaments. The relevant metric is the uniaxial tensile bond strength at age, and its sensitivity to layer-cycle time, surface drying, and substrate roughness. The microstructural driver is the maximum local porosity in the interlayer band, not the average porosity. Bond strength generally peaks at short layer-cycle times and degrades as the substrate stiffens and dries before the new layer can wet it (Chen et al., 2020; Marchment et al., 2019).
Dimensional stability is the property of the printed structure to retain its as-printed geometry as drying, hydration shrinkage, and elastic deformation under self-weight act on it. The metrics are total free shrinkage, restrained shrinkage cracking risk, and elastic settlement under self-weight. Dimensional stability couples back into the mix design through paste fraction, water content, the choice of viscosity-modifying admixture (VMA), and SCM substitution.
The five regimes are coupled. Increasing static yield stress to improve buildability typically reduces bond strength at long layer-cycle times because the substrate stiffens before the new layer can wet it. Increasing pumpability by adding water or superplasticizer lowers buildability. Tightening the granular skeleton to reduce shrinkage often raises plastic viscosity beyond pump capacity. The printability window is the simultaneous solution to all five constraints, and any tool that claims to predict it must respect that the axes are not separable.
4. Particle packing as the foundation: the Modified Andreasen-Andersen model
Below the rheology there is geometry. The dry skeleton of a 3DCP mortar, cement plus SCMs plus fines plus sand, must pack densely enough that paste demand is low, water demand is low, and the resulting suspension is rheologically tractable. Too loose a packing and water and superplasticizer demand explode, taking embodied carbon and shrinkage with them. Too tight and the mix becomes harsh, extrudability suffers, and surface finish degrades.
The Modified Andreasen and Andersen (MAA) packing model is the literal backbone of CEMFORGE’s gradation theory. In its standard form, MAA expresses the target cumulative passing fraction at particle diameter D as a function of the upper and lower size bounds and a single distribution exponent q that controls the slope (Funk and Dinger, 1994; Brouwers, 2006). Funk and Dinger’s monograph remains the canonical reference, and we cite it as the theoretical foundation rather than as a historical footnote. The exponent q is the dial. Lower q shifts cumulative mass into the fines, which raises specific surface area, raises water demand, and stabilizes the fresh state against bleeding, segregation, and filter pressing. Higher q opens the gradation, lowering surface area and water demand but reducing the fines fraction available for early structuration. For conventional self-compacting concrete the optimum sits in one part of the range; for 3DCP mortars, where pumpability and extrudability constrain the upper end of the gradation and where shape stability rewards a higher fines fraction, the optimum sits lower (Weng et al., 2018; Hüsken and Brouwers, 2008).
Why does q matter so much in 3DCP specifically? At fixed bounds, a lower q increases the specific surface area of the dry skeleton, which raises water demand at constant flow but also stabilizes the fresh state. It also lowers the spacing between solid particles inside the paste, which raises static yield stress and the structural buildup rate. The price is paid in superplasticizer demand and in embodied carbon if the fines are clinker. Replacing clinker fines with limestone filler, fly ash, or calcined clay recovers the carbon budget but changes the surface chemistry of the paste, with second-order effects on hydration kinetics and on adsorption of polycarboxylate ether superplasticizers.
The point is that particle packing is upstream of rheology. Two formulations with the same chemistry but different gradations can sit in entirely different rheological regimes. CEMFORGE’s formulation engine validates candidate mixes against the MAA target curve as a hard physical gate before any rheological prediction is made. If a candidate skeleton cannot physically close volumetrically against the target distribution at the user’s bounds, the engine rejects it rather than passing it to a downstream model that has no information about gradation closure.
A subtler point is that “packing” in 3DCP is not just a geometric optimum, it is a manufacturing constraint. The fines fraction that the MAA curve calls for must be physically procurable in the local market at consistent quality. Wisconsin limestone fines and Texas limestone fines are not interchangeable as MAA inputs because their D50, D90, particle shape, and surface chemistry differ. The same is true for Class F fly ashes from different sources, calcined clays from different deposits, and slags from different mills (Lowke et al., 2018). This is why open data standards matter: a packing model trained or tuned on one dataset will fail predictably on another set of locally sourced materials unless the metadata for each constituent is captured.
It is also worth saying that the MAA target curve is necessary but not sufficient. Two skeletons that share the same cumulative passing curve can pack very differently if the constituent particles have different shape factors, different surface roughness, or different aspect ratios. Calcined clay, in particular, often plates rather than packs as a sphere of equivalent diameter. Fly ash spheres pack closer to ideal MAA targets than crushed limestone fines of the same nominal D50. In production, the gradation match against MAA should be paired with at least a measured void index of the dry blend (loose and tapped bulk densities are sufficient first-pass instrumentation) so that the engineer is not fooled by a curve that looks correct but a skeleton that is not actually closed.
5. Static versus dynamic yield stress
The single most useful mental model in 3DCP rheology is the distinction between static and dynamic yield stress. They look similar in equations and they are routinely confused in conversation, but they capture different physics and they answer different engineering questions.
Dynamic yield stress is the stress at which a mortar that is already flowing continues to flow. It is the intercept of the descending branch of the flow curve in a controlled-shear-rate rheometer, fit to a Bingham or Herschel-Bulkley model. Operationally it is the stress that the pump and nozzle must overcome to keep the filament moving. The relevant test geometries are the vane-in-cup rheometer and the parallel-plate rheometer. Slump-cone tests do not capture dynamic yield stress reliably for mortars at this consistency (Roussel, 2018; Tay et al., 2019).
Static yield stress is the stress at which a mortar at rest first begins to flow. It is the stress that a deposited layer must withstand without yielding under the load of the layers above it. Operationally it determines buildability. Static yield stress is a function of rest time, and it grows under three coupled mechanisms: colloidal flocculation of cement particles, early ettringite and calcium silicate hydrate (C-S-H) nucleation, and physical structuration of the suspension (Roussel, 2018; Perrot et al., 2016). The standard engineering approximation models static yield stress as a baseline value plus a structural buildup rate multiplied by rest time. The metric is measured by repeated vane tests at increasing rest times, by stress-growth tests in a rheometer, or by penetration tests calibrated to vane data.
Both yield stresses matter, and they matter independently. A mortar can have a high dynamic yield stress (hard to extrude, low pumpability margin) and a low structural buildup rate (poor buildability), and there is no single test that reveals the combination. A vane rheometer captures static yield stress over time if run at very low shear and increasing rest times. The same instrument captures dynamic yield stress if run as a flow curve at imposed shear rate. A slump-cone test captures, at best, an integral of static yield stress at rest. A flow-table test captures something closer to dynamic yield stress under impact loading, but with poor reproducibility for stiff mortars. The 3DCP rheology literature converges on vane rheometry and stress-growth testing as the minimum credible instrumentation for mix qualification today (Mechtcherine et al., 2020; De Schutter et al., 2018), and we adopt that view where the field has converged on it while staying open to refinements as inline and image-based methods mature.
The buildability prediction that engineers actually need is not the baseline static yield stress alone but the static yield stress at the specific layer-cycle time imposed by the printer. The mortar at the bottom of a fresh column must support the cumulative overburden of the layers above it before the printer reaches the topmost fresh material. The mortar’s structural buildup must outrun that overburden. Whether it does is set by the chemistry, by the SCM choice, by the admixture package, and by ambient conditions, and the arithmetic is unforgiving. The input stresses have to be measured, not assumed.
6. Why machine learning helps: a digital twin of cementitious-mortar behavior
The argument for machine learning in 3DCP mix design is not that it replaces materials science. It is that the materials-science problem is high-dimensional, the response surface is noisy, and the closed-form theories cover only fragments of it. A modern printable mortar formulation has many independent levers: cement type, SCM identity and proportion, filler identity and proportion, sand grading, water-to-binder ratio, superplasticizer chemistry and dosage, VMA chemistry and dosage, accelerator if any, and ambient conditions. Each of those couples to multiple regimes (pumpability, buildability, layer adhesion, hardened strength, embodied carbon) through nonlinear and often non-monotone relationships. Empirical models trained on a curated dataset of validated print specimens capture interactions that a pure first-principles model cannot.
CEMFORGE’s machine-learning architecture is best understood not as a property-by-property regressor but as a digital twin of cementitious-mortar behavior. The twin spans four families of inputs simultaneously: chemistry (oxide composition, Bogue phases, SCM identity and substitution rate, admixture chemistry), particle-size distribution (the MAA-conformant gradation discussed in Section 4, plus shape and surface metadata for each constituent), process variables (mixing energy and duration, pump curve, nozzle geometry, layer-cycle time, print speed), and environmental conditions (temperature, humidity, substrate state). Those four domains are exactly the columns the Open3DCP schema captures, by design: the schema is the data contract for the twin, and any specimen entered into the loop carries all four families of metadata or is rejected for re-training purposes.
The current generation of the platform is a stacked machine-learning ensemble keyed to those input families and gated by physics. The trajectory is toward an artificial neural network that captures the cross-domain couplings end-to-end on a sufficient training corpus. We say ML now, ANN trajectory plainly, without overclaiming the latter. Both architectures live inside the same digital-twin framing because the input contract is the same: chemistry plus particle-size distribution plus process plus environment, schema-aligned, with provenance.
This stance has a precedent. QuesTek Innovations operated a closed proprietary ICME platform branded Materials by Design through which alloy design briefs were converted into qualified compositions for aerospace and defense (Olson, 1997, 1998). The platform’s defensibility came not from any single model but from the discipline of an end-to-end ICME loop with a controlled data contract. CEMFORGE follows the same posture for cementitious systems: the loop is the asset, the schema is the connective tissue, and any individual model inside the loop is replaceable as the data corpus grows.
Single-model approaches fail predictably on this problem. An XGBoost-only or neural-net-only predictor, trained on the available 3DCP literature plus internal data, performs well on within-distribution formulations and silently fails on out-of-distribution ones. Out-of-distribution failure modes in 3DCP include a novel SCM substitution rate the training set never saw, a sand grading outside the trained envelope, a superplasticizer chemistry the model has no examples of, or a regional limestone fines source with a different particle morphology than the training data. Single-model predictors will return a printability score for any of those inputs, and the score will be wrong without flagging itself.
The defensive posture is multi-model ensembles plus physics-based gating. CEMFORGE combines physical particle-packing validation against the MAA target, a property-by-property model selection where models are trained only where the validated training data meaningfully covers the target property, and explicit acceptance thresholds drawn from current widely-used experimental conventions (including, but not limited to, RILEM-style reporting). The system gates rheology predictions on whether the candidate gradation can physically close volumetrically, gates buildability predictions on whether the predicted structural buildup trajectory stays inside the trained envelope, and gates hardened-strength predictions on whether the predicted hydration trajectory has training support. Where any gate trips, the system either declines to predict or returns a flagged uncertainty band.
A final, blunter point about ML in 3DCP. The main quality bottleneck in 3DCP machine learning today is not algorithm choice. It is dataset quality and metadata completeness. A model trained on a large but inconsistently labeled aggregation of literature mixes will silently encode the conventions of whichever lab contributed the most data. A model trained on a smaller but rigorously schema-aligned dataset, with full metadata on cement chemistry, gradation, admixture identity, mixing protocol, rheology test geometry, and printer configuration, will outperform the larger model on out-of-distribution prediction. This is the argument for an open data standard like Open3DCP (https://open3dcp.org, DOI 10.5281/zenodo.19647471), which provides a flat schema for 3DCP mix-design data with first-class fields for chemistry, gradation, process parameters, rheology, environmental conditions, and hardened properties. The Open3DCP schema is the data contract for CEMFORGE’s digital twin, and the twin is only as strong as the schema’s coverage of the four ICME domains.
7. The print window
The print window is the interval, in time after water contact, during which a mortar simultaneously satisfies pumpability, extrudability, and buildability constraints. Outside that window, on either side, a print will fail. On the early end, the window opens once the mortar has been mixed long enough for superplasticizer to disperse cement agglomerates, for VMA to hydrate, and for the dynamic yield stress to settle into a stable value. On the late end, the window closes when the mortar is no longer pumpable, typically because the static yield stress has grown to where pump pressure exceeds available capacity, or because extrudability has degraded to where filaments tear at the nozzle. The closure time depends on cement chemistry, retarder dosage if any, VMA, and ambient temperature. Open times for printable mortars span a wide range; chemically retarded mortars extend the window, and accelerated set-on-demand mortars compress it (Reiter et al., 2018).
The print window must be a predicted range, not a single number, because the field operator cannot afford to discover its boundaries by failure. CEMFORGE’s predictions, and the predictions of any credible mix-design tool for 3DCP, must therefore output an open-time floor, an open-time ceiling, and the rate of change of static yield stress across the interval, with quantified uncertainty.
A subtler issue: the print window is not stationary across an actual job. Ambient temperature drifts. The aggregate stockpile heats in the sun. Mixer wear changes shear input. If batch-to-batch reproducibility of static yield stress is loose, the print window shrinks effectively even when the nominal numbers look fine. RILEM TC 304-ADC’s interlaboratory work, among other available reference points, has documented that batch-to-batch variability is larger than most lab-scale studies acknowledge and is the dominant scatter source in cross-lab comparisons (Robens-Radermacher et al., 2025; Mechtcherine et al., 2020). Production sites have to budget for it, and a digital twin that ignores environmental and batch-to-batch variability is a twin that will work in the lab and fail on site.
8. Embodied carbon and SCM substitution
Mix design today is mix design under a carbon ceiling. The reference benchmark is ordinary Portland cement, whose clinker carries a large embodied-carbon burden per unit mass; substituting a meaningful fraction of clinker with supplementary cementitious materials is no longer optional in most jurisdictions and most procurement specifications (Scrivener et al., 2018). The question is which SCMs and how they affect the printability window.
Fly ash (Class F) reduces clinker demand at low cost in most North American markets, slows early hydration, and lowers the structural buildup rate, which extends open time but reduces buildability per layer. The fly-ash particle is roughly spherical, which improves pumpability through a ball-bearing effect but reduces the surface area available to seed early structuration. In a 3DCP context, fly-ash-rich blends typically have to be paired with a thixotropy-enhancing VMA or a small accelerator dose to recover buildability.
Slag (GGBS) is widely used in conventional concrete and provides comparable carbon savings to fly ash. In 3DCP, slag tends to extend open time beyond what most printers can use effectively, and the late strength gain is helpful for hardened performance but does not help buildability. Slag-rich blends in 3DCP are most effective when paired with a small calcined-clay or silica-fume fines fraction to recover early structuration (Adu-Amankwah et al., 2017).
Calcined clay (typically metakaolin-rich kaolinite calcined in the canonical temperature range) is the high-leverage SCM for 3DCP. It raises both static and dynamic yield stress materially, raises the structural buildup rate, and develops strength faster than fly ash or slag. The mechanism is partly geometric (calcined clays are platy and irregular, raising water demand and shear stiffness at modest dosage) and partly chemical (metakaolin reacts with portlandite to form additional C-A-S-H). The cost is high water demand and high superplasticizer demand, plus regional supply variability (Chen et al., 2020).
Limestone fillers (ground to a fine D50) are increasingly the workhorse fines in low-carbon 3DCP mortars. Limestone is dilution rather than substitution at low dosage, but it improves packing, marginally accelerates clinker hydration via nucleation effects, and is effectively zero embodied carbon at the inert end. The combined LC3 system (Limestone Calcined Clay Cement, roughly half clinker, with a smaller calcined-clay fraction, a limestone fraction, and a gypsum fraction) is the canonical low-carbon binder for 3DCP and has been the subject of substantial dedicated research (Scrivener et al., 2018; Adu-Amankwah et al., 2017).
The carbon-buildability tradeoff is real and not always monotone. Replacing clinker with fly ash drops embodied carbon and lowers the structural buildup rate, often forcing additional admixture dosing that has its own carbon cost. Replacing clinker with calcined clay drops embodied carbon and raises the structural buildup rate, but raises water demand, which can raise paste fraction, which raises shrinkage and embodied carbon per cubic metre of placed mortar. The optimization is multi-objective, the constraints are coupled, and the right answer depends on local material chemistry, local carbon pricing, and the specific printability targets of the project. CEMFORGE’s formulation engine treats embodied carbon as a first-class output alongside printability metrics rather than a downstream report, which is consistent with where industry briefs are heading: most current client mix-design briefs we see specify both an embodied-carbon ceiling and a hardened compressive-strength floor as joint constraints.
9. From dataset to deployment
A model is only as good as its training data, and a 3DCP model is only as good as the match between its training data and the materials a contractor will actually use on site. This is more demanding for 3DCP than for conventional concrete, because the printability window is narrow and because regional material chemistry varies enough to push a candidate mix outside the trained envelope.
The instructive failure mode is an academic dataset trained on European cements and European calcined clays, then deployed in a North American project with Type IL portland-limestone cement, Class F fly ash, and locally graded river sand. The model’s predictions are not wrong in the sense of being internally inconsistent. They are wrong in the sense of being out of distribution. The fines fraction the model expects from the binder is actually present, but the chemistry is different. The hydration curve is not what the training data assumes. Buildability prediction drifts.
This is why dataset transparency matters. Open3DCP (https://open3dcp.org, DOI 10.5281/zenodo.19647471) is a flat schema for 3DCP mix-design data with explicit columns for cement chemistry (oxides and Bogue phases where available), SCM identity and source, gradation by D10, D50, D90, and span, admixture chemistry and dosage, mixing protocol, rheology test geometry, printer configuration, layer-cycle time, ambient conditions, and hardened mechanical properties. The schema is licensed Apache 2.0, is compatible with current widely-used conventions including ASTM, RILEM, and NIST references (without committing to any single methodology), and is the data format CEMFORGE uses internally for training-data audit and model-coverage diagnostics. The prediction stack is deliberately open to newer measurement protocols, including in-line process telemetry, computer-vision-based bond inspection, and emerging digital-twin instrumentation, which are first-class features as the field develops them. It is also the contract that allows the digital twin in Section 6 to ingest data from independent labs without silently importing whichever lab’s conventions came pre-baked.
The practical consequence is that a 3DCP mix-design tool should tell the user not only what its prediction is, but how well its training data covers the user’s specific materials. CEMFORGE’s intended UX is to surface a training-coverage indicator alongside each prediction, so that a contractor running a regional Type IL plus regional Class F plus locally crushed dolomitic limestone fines knows whether the model is interpolating or extrapolating. Where coverage is thin, the recommended workflow is a small confirmatory bench print rather than direct production deployment. Local materials matter. Wisconsin limestone is not Texas limestone, both are valid 3DCP feedstock, but the model has to know which one is in the mixer.
This dataset-transparency posture is also why CEMFORGE is paired with M3-CRETE (https://m3-crete.com), the printer that consumes these mixes. The pairing closes the ICME loop in hardware: CEMFORGE proposes a candidate formulation under the ICME framing of Section 2, M3-CRETE prints a small-scale verification specimen with known nozzle, layer, and ambient parameters, and the resulting fresh-state and hardened data are captured back into the Open3DCP schema for re-training. The point is not vertical integration for its own sake; it is to keep the training data and the deployment context tied together, which is exactly the leak that opens on out-of-distribution mixes.
10. CEMFORGE in practice
If you are designing a 3DCP mix today, CEMFORGE generates particle-packing validation against an MAA target, fresh-state rheology predictions, printability-window estimates, embodied-carbon outputs, and hardened compressive-strength predictions from your local materials and your target performance. The platform combines a Modified Andreasen-Andersen packing engine, a stacked machine-learning ensemble that captures chemistry, particle-size distribution, process variables, and environmental conditions as a single digital twin (with an ANN-based successor on the trajectory), and physics-based out-of-distribution gates that flag low-coverage inputs rather than silently extrapolate. The intent is decision support for the engineering team, not a replacement for the engineering team, and not a black box. Sunnyday Technologies (https://sunn3d.com), the company behind CEMFORGE, also maintains the M3-CRETE 3DCP gantry platform (https://m3-crete.com), the CADCLAW (https://cadclaw.io) open-source CAD validation framework that gates the M3-CRETE hardware itself before fabrication, and the Open3DCP open data standard (https://open3dcp.org), so the predictions, the printer that produces verification specimens, the engineering tooling that validates the printer, and the data schema that captures the result all live inside one engineered ICME loop. Try CEMFORGE at cemforge.ai.
References
Adu-Amankwah, S., Zajac, M., Stabler, C., Lothenbach, B., Black, L. (2017). Influence of limestone on the hydration of ternary slag cements. Cement and Concrete Research, 100, 96-109. DOI: 10.1016/j.cemconres.2017.05.013
Brouwers, H. J. H. (2006). Particle-size distribution and packing fraction of geometric random packings. Physical Review E, 74(3), 031309. DOI: 10.1103/PhysRevE.74.031309
Buswell, R. A., Leal de Silva, W. R., Jones, S. Z., Dirrenberger, J. (2018). 3D printing using concrete extrusion: A roadmap for research. Cement and Concrete Research, 112, 37-49. DOI: 10.1016/j.cemconres.2018.05.006
Chen, Y., Jansen, K., Zhang, H., Romero Rodriguez, C., Gan, Y., Copuroglu, O., Schlangen, E. (2020). Effect of printing parameters on interlayer bond strength of 3D printed limestone-calcined clay-based cementitious materials: An experimental and numerical study. Construction and Building Materials, 262, 120094. DOI: 10.1016/j.conbuildmat.2020.120094
De Schutter, G., Lesage, K., Mechtcherine, V., Nerella, V. N., Habert, G., Agusti-Juan, I. (2018). Vision of 3D printing with concrete: Technical, economic and environmental potentials. Cement and Concrete Research, 112, 25-36. DOI: 10.1016/j.cemconres.2018.06.001
Funk, J. E., Dinger, D. R. (1994). Predictive Process Control of Crowded Particulate Suspensions Applied to Ceramic Manufacturing. Springer, Boston, MA. ISBN: 978-1-4615-3118-0. (Theoretical foundation of the Modified Andreasen-Andersen packing model.)
Hüsken, G., Brouwers, H. J. H. (2008). A new mix design concept for earth-moist concrete: A theoretical and experimental study. Cement and Concrete Research, 38(10), 1246-1259. DOI: 10.1016/j.cemconres.2008.04.002
Lowke, D., Dini, E., Perrot, A., Weger, D., Gehlen, C., Dillenburger, B. (2018). Particle-bed 3D printing in concrete construction, possibilities and challenges. Cement and Concrete Research, 112, 50-65. DOI: 10.1016/j.cemconres.2018.05.018
Marchment, T., Sanjayan, J., Xia, M. (2019). Method of enhancing interlayer bond strength in construction scale 3D printing with mortar by effective bond area amplification. Materials and Design, 169, 107684. DOI: 10.1016/j.matdes.2019.107684
Mechtcherine, V., Bos, F. P., Perrot, A., Leal da Silva, W. R., Nerella, V. N., Fataei, S., Wolfs, R. J. M., Sonebi, M., Roussel, N. (2020). Extrusion-based additive manufacturing with cement-based materials, production steps, processes, and their underlying physics: A review. Cement and Concrete Research, 132, 106037. DOI: 10.1016/j.cemconres.2020.106037
Olson, G. B. (1997). Computational design of hierarchically structured materials. Science, 277(5330), 1237-1242. DOI: 10.1126/science.277.5330.1237
Olson, G. B. (1998). Systems design of hierarchically structured materials: Advanced steels. Journal of Computer-Aided Materials Design, 4(3), 143-156. DOI: 10.1023/A:1008670319664
Perrot, A., Rangeard, D., Pierre, A. (2016). Structural built-up of cement-based materials used for 3D-printing extrusion techniques. Materials and Structures, 49(4), 1213-1220. DOI: 10.1617/s11527-015-0571-0
Reiter, L., Wangler, T., Roussel, N., Flatt, R. J. (2018). The role of early age structural build-up in digital fabrication with concrete. Cement and Concrete Research, 112, 86-95. DOI: 10.1016/j.cemconres.2018.05.011
Robens-Radermacher, A., Kujath, C., Bos, F., Mechtcherine, V., Unger, J. F. (2025). Mechanical properties of 3D printed concrete: a RILEM TC 304-ADC interlaboratory study, design and implementation of a database system for querying, sharing, and analyzing experimental data. Materials and Structures, 58(5), 184. DOI: 10.1617/s11527-025-02650-9
Roussel, N. (2018). Rheological requirements for printable concretes. Cement and Concrete Research, 112, 76-85. DOI: 10.1016/j.cemconres.2018.04.005
Roussel, N., Lowke, D. (Eds.) (2022). Digital Fabrication with Cement-Based Materials, State-of-the-Art Report of the RILEM TC 276-DFC. Springer International Publishing. DOI: 10.1007/978-3-030-90535-4
Scrivener, K. L., John, V. M., Gartner, E. M. (2018). Eco-efficient cements: Potential, economically viable solutions for a low-CO2 cement-based materials industry. Cement and Concrete Research, 114, 2-26. DOI: 10.1016/j.cemconres.2018.03.015
Secrieru, E., Mohamed, W., Fataei, S., Mechtcherine, V. (2020). Assessment and prediction of concrete flow and pumping pressure in pipeline. Cement and Concrete Composites, 107, 103495. DOI: 10.1016/j.cemconcomp.2019.103495
Sunnyday Technologies (2025). Open3DCP, an open data schema for 3D concrete printing mix-design records (V1.5.0). Zenodo. DOI: 10.5281/zenodo.19647471
Tay, Y. W. D., Qian, Y., Tan, M. J. (2019). Printability region for 3D concrete printing using slump and slump flow test. Composites Part B, Engineering, 174, 106968. DOI: 10.1016/j.compositesb.2019.106968
Wangler, T., Lloret, E., Reiter, L., Hack, N., Gramazio, F., Kohler, M., Bernhard, M., Dillenburger, B., Buchli, J., Roussel, N., Flatt, R. J. (2016). Digital concrete: Opportunities and challenges. RILEM Technical Letters, 1, 67-75. DOI: 10.21809/rilemtechlett.2016.16
Weng, Y., Li, M., Tan, M. J., Qian, S. (2018). Design 3D printing cementitious materials via Fuller Thompson theory and Marson-Percy model. Construction and Building Materials, 163, 600-610. DOI: 10.1016/j.conbuildmat.2017.12.112