The motivation behind the project lies in a persistent bottleneck in materials research. Solution-based printing, the family of techniques that includes inkjet, aerosol jet, gravure and screen printing, underpins modern work on flexible electronics, chemical sensors and printed diagnostics, all of which depend on depositing inks and functional fluids in exact locations on diverse substrates. Commercial tools such as the Fujifilm Dimatix Materials Printer DMP-2850 handle this task admirably, but their price places them beyond many academic groups. Meanwhile, the drive toward self-driving laboratories, where robots, liquid handlers and inline instruments work in concert to accelerate discovery, has exposed another cost barrier: automation platforms like the Opentrons OT-2 robot start at more than 15,000 dollars before a single experiment runs. The CRISP team set out to demonstrate that a frugal twin could be assembled from commodity hardware without sacrificing meaningful performance.

The engineering insight at the heart of the project is that a fused deposition modeling printer already contains nearly everything a liquid-dispensing robot needs. The Ender-3 provides robust three-axis motion control, a stepper motor for extrusion, an Arduino-class motherboard running Marlin firmware, and a mature G-code ecosystem. The researchers removed the hot end, heating block and filament spool holder, then mounted a syringe pump assembly on the printer’s upper frame. That pump is deliberately simple: a stainless-steel trapezoidal lead screw driven by the printer’s stock extruder stepper motor, a brass flanged nut riding on a 3D-printed gantry bracket, and two mounts that clamp a standard 20 milliliter Luer-lock syringe in place. When the motor turns, the threaded rod advances and the gantry bracket pushes the syringe plunger with kinematic regularity, converting rotational steps into volumetric dispensing.

Ink reaches the substrate through a length of fluorinated ethylene propylene tubing, roughly 75 centimeters long with a one-sixteenth-inch inner diameter, which press-fits into a nozzle made from a disposable polypropylene pipette tip. The choice of pipette tips as nozzles is a small piece of design genius: at 10 to 20 cents apiece, they are effectively consumables, and a clogged nozzle is solved by pulling the tip off and pressing on a new one rather than performing delicate surgery on an expensive printhead. The tips used in the demonstration have orifice diameters between roughly 0.35 and 0.50 millimeters, but because deposition is volumetric rather than nozzle-limited, the final droplet size on porous paper is governed mainly by the dispensed volume and how the fluid wicks into the fibers.

Because the syringe pump borrows the printer’s native extruder motor, it is programmed with ordinary G-code, the same command language hobbyists use to print plastic. This gives CRISP a surprisingly sophisticated feature set for its price: the number of programmed dispensing steps is limited only by how many G-code commands a user writes, which allows multi-step routines comparable to those of programmable laboratory pumps that cost several thousand dollars. For calibration, the team primed the tubing through the printer’s own interface, loaded a G-code file onto a microSD card, and let the machine execute its run from file. The comparison point is stark: a Chemyx 4000X programmable syringe pump runs about 4,400 dollars, while CRISP’s pumping hardware costs a fraction of that and integrates directly into a motion platform.

Validation began with a deliberately humble test fluid, a blue dye dissolved in water, chosen so that printing parameters could be optimized without the expense and opacity of metal nanoparticle inks. The team printed repeated five-by-five arrays of droplets on cardstock with dots spaced 20 millimeters apart and a programmed volume of 0.026 milliliters per drop. Quantitative image analysis using the open-source computer vision library OpenCV revealed a mean droplet circularity of 0.87 with a standard deviation of only 0.02, a coefficient of variation of 2.30 percent that indicates highly uniform droplet morphology. Positional accuracy proved equally respectable: the mean deviation between programmed and actual droplet centers was 0.65 millimeters, with a worst case of 1.25 millimeters, and the error distribution was random rather than systematic, confirming that the added syringe hardware introduces no mechanical drift into the printer’s kinematics.

Weighing the droplets provided an independent check on volumetric fidelity. Across five trials in which the system dispensed ten droplets into a tared dish, the average total mass was 265.9 milligrams, implying 26.59 microliters per drop against a kinematic prediction of 26.0 microliters, a percent error of just 2.3 with a coefficient of variation of 1.6 percent. Deposited on porous cardstock, the drops spread into hemispheres averaging about 4.6 millimeters in diameter, exactly matching volumetric expectations. The researchers note that the practical minimum droplet volume for this configuration is roughly 3 to 5 microliters, bounded by the stepper motor’s step resolution and fluid surface tension, though switching to a smaller syringe barrel would stretch that limit further by amplifying plunger travel per dispensed volume.

The team is candid about the platform’s boundaries. The nozzle holder as built maintains a physical gap of about 7 millimeters between tip and substrate, which is ideal for falling droplets but prevents the close tip-to-surface contact needed for a stable meniscus during continuous line printing; attempts to write continuous features with aqueous dye produced lines that thinned and broke. Paper remains the sweet spot because its fibrous structure absorbs liquid quickly, pinning particles before uneven drying can occur, whereas non-porous glass slides exhibited anomalous drying behavior related to well-known evaporative defects such as the coffee ring effect. The dead volume of the delivery tubing, roughly 1.5 milliliters, also matters when working with expensive reagents, and the authors recommend shorter, stiffer tubing and a heated print bed to tame environmental drying variability.

What makes CRISP more than a clever hack is its framing within the growing frugal-twin movement in chemistry and materials science, where low-cost replicas of commercial instruments are validated against their expensive counterparts. Structurally, the design resembles prior Ender-3 liquid-handling conversions, but it differs in relying entirely on stock printer hardware and 3D-printed brackets rather than an external syringe pump, keeping the added cost below 200 dollars. With results showing dispensing circularity of 98.2 percent and relative positional accuracy of 96.8 percent, the authors argue that stepper-driven commodity hardware offers sufficient repeatability for routine patterning tasks. Their ultimate goal is the deposition of metal nanoparticle inks for paper-based sensors, and with every CAD file, STL, G-code script and dataset published under open licenses on the Open Science Framework, CRISP invites laboratories worldwide to print droplets, not just plastic, for the price of a good dinner out.

Subject of Research: A low-cost 3D printer-based platform for precision solution printing on paper-based substrates

Article Title: CRISP: A 3D printer-based platform for precision solution printing on paper-based substrates

Article References: Lutfiyev, I., McCoy, S. A., Star, R., Giordano, A. N., Rist, B., Baldwin, L. A., & Rao, R. (2026). CRISP: A 3D printer-based platform for precision solution printing on paper-based substrates. HardwareX, 28, Article e00842. https://doi.org/10.1016/j.ohx.2026.e00842

Image Credits: AI Generated

DOI: 10.1016/j.ohx.2026.e00842

Keywords: CRISP, 3D printing, syringe pump, solution printing, paper-based substrates, Ender-3, open-source hardware, G-code, droplet deposition, frugal science, flexible sensors, HardwareX

The study, published in Results in Engineering by Kazuki Kaneko and Zongze Li, addresses a fundamental blind spot in digital manufacturing. Computer-aided manufacturing (CAM) software can simulate geometry and generate tool paths from a CAD model, but it cannot simulate what happens inside the machine’s controller. The G-code is filtered by the CNC interpolator, feedforward controllers and feedback control loops before it ever reaches the servomotors, and these proprietary algorithms are closely guarded black boxes maintained by machine tool builders. Physical modeling of this chain is extraordinarily difficult precisely because the control algorithms are secret and the dynamics are deeply nonlinear. The result is that manufacturers often discover machining errors only after cutting — or resort to time-consuming air runs in which the machine executes the program without a workpiece just to observe its behavior.

The authors reframed the problem as one of sequence translation. From a data-processing perspective, G-code execution is a stream of discrete commands fed row by row into a black-box system that emits corresponding sequences of tool-path and cycle time errors. Predicting those errors for a given program is therefore fundamentally a sequence-to-sequence problem — exactly the kind of task at which Transformers excel. Unlike recurrent networks or long short-term memory (LSTM) models, which process sequences step by step, a Transformer uses multi-head self-attention to derive global relational weights across all positions simultaneously. This allows it to capture long-range temporal dependencies, such as how a sharp corner several blocks ahead causes the servo system to decelerate and how a large acceleration from a previous segment continues to influence subsequent commands as the transient response gradually decays.

The researchers drew a striking analogy between the Transformer’s internal machinery and the physics of a CNC controller. In their framing, the query represents the motion state of the current G-code command, while the keys and values carry error features of the surrounding tool path; scaled dot-product attention then assigns weights to nearby and distant blocks, mirroring the feedforward and feedback structure of a real servo loop. Position-wise feedforward networks introduce the nonlinear modeling capacity needed to emulate local servo behaviors such as PID control and friction, while residual connections — which pass inputs directly forward — resemble a closed-loop feedback system that focuses on predicting small dynamic deviations rather than reconstructing the entire tool path. Layer normalization, meanwhile, stabilizes the wildly different physical scales of position increments and feed rates, functioning much like a gain stabilizer in a servo drive.

Training such a model requires data that spans the machine’s full dynamic repertoire, so the team avoided reliance on specific workpiece geometries, which would limit generalization. Instead, they semi-randomly generated ten sets of training G-code combining local blending segments, where the path passes directly through control points, and global blending segments, where B-spline interpolation smooths the trajectory at constant feed rate. Displacements and feed rates were drawn randomly from defined ranges, producing tool paths with varied travel ranges and frequent direction changes. Each program was then executed on a real three-axis ROBODRILL machine equipped with a Fanuc Series 31i controller, and actual tool-path data were captured at a sampling frequency of 1 kHz using Fanuc’s SERVO Guide monitoring software. The training targets were thus experimentally measured servo responses rather than virtual simulations, grounding the digital twin in physical reality.

The model itself is an encoder-only Transformer with a model dimension of 512, eight layers and four attention heads, trained for 2000 epochs on a workstation with an RTX 4090 GPU. Input sequences consisted of nominal positional increments along the three feed axes plus the commanded feed rate, segmented into windows of 32 control points using a causal sliding window. The output comprised the three-dimensional tool-path error and the cumulative cycle time error at each point. Notably, the authors found that longer input sequences converged more slowly without improving final accuracy — a result that aligns neatly with servo physics, since the transient influence of any single command naturally fades over time. When compared against a Bi-LSTM model of equal dimensions trained on identical data, the Transformer converged in roughly 500 epochs, three times faster than its recurrent rival.

Validation was carried out on six G-code programs derived from two physical specimens: a 25-millimeter-radius hemispherical part machined along morphing spiral and bidirectional paths, and a complex artifact geometry, each run at nominal feed rates of 1000 and 3000 millimeters per minute. The results were decisive. Both models avoided large systematic bias, with mean prediction deviations generally within ±1 micrometer or millisecond, but the Transformer’s uncertainty bands were roughly 1.5 to 2 times tighter than the Bi-LSTM’s across all error components — and in some conditions nearly four times tighter. Its error distributions clustered markedly more sharply around zero, and while the Bi-LSTM’s cycle time root-mean-square error ballooned to about 8 milliseconds on the most demanding program against a mean error of 2.5 milliseconds, the Transformer suppressed such outliers, keeping RMSE close to MAE.

The comparison also revealed how differently the two architectures degrade under stress. Higher feed rates inherently reduce the accuracy of the controller’s feedforward and feedback loops, and the Bi-LSTM’s errors doubled or even tripled when the feed rate rose from 1000 to 3000 millimeters per minute. The Transformer, by contrast, showed far more graceful degradation, keeping mean errors below 2 micrometers or milliseconds even on the most challenging high-speed program. The authors attribute the residual errors that remain — small but nonzero — to highly random mechanical nonlinearities such as micro-friction variations and backlash during axis reversals, along with the quantization limits of encoder feedback at the 1 kHz sampling rate.

Perhaps the most industrially compelling result is speed. Predicting errors for an entire validation program took the Transformer less than 3.5 seconds, regardless of program complexity, whereas air-running the same programs on the machine took anywhere from roughly 38 to 202 seconds. Because prediction time depends only on G-code length, the trained model can instantly screen massive production programs offline, eliminating the need for preliminary air runs and sharpening both production scheduling and machining efficiency. The approach does not generalize across machines — each machine tool has unique controller and mechanical characteristics — but the authors emphasize that adapting to a new machine requires only executing the training G-codes, acquiring a fresh dataset and retraining, after which the model serves as a customized digital surrogate for that specific machine.

The work points toward a future in which every machine tool carries its own learned digital twin, capable of vetting programs before metal is ever cut. The team’s next steps include extending the model to account for cutting force-induced deflections of tool and workpiece, which superimpose on servo tracking errors in real machining; probing how specific G-code features such as control point density affect prediction performance; scaling the sequence-to-sequence methodology to five-axis simultaneous machining; and developing real-time servo signal acquisition for online error prediction during actual cutting. They also plan to build open-architecture experimental CNC systems, since proprietary commercial controllers block the integration of theoretical control equations needed for hybrid physical-digital modeling. If those ambitions pan out, the humble G-code file may one day arrive at the shop floor already annotated with a complete forecast of its own imperfections.

Subject of Research: Transformer-based deep learning prediction of tool-path and cycle time errors in CNC machining from G-code

Article Title: Prediction of tool-path and cycle time errors of CNC machining using a transformer-based deep learning model

Article References: Kaneko, K., & Li, Z. (2026). Prediction of tool-path and cycle time errors of CNC machining using a transformer-based deep learning model. Results in Engineering, 32, Article 112872. https://doi.org/10.1016/j.rineng.2026.112872

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.112872

Keywords: CNC machining, Transformer model, deep learning, digital twin, tool-path error, cycle time prediction, G-code, servo dynamics, Bi-LSTM comparison, precision manufacturing, sequence-to-sequence learning, machine tools