Printer: Advanced Deposition & Control

• Printer is a system that spatially deposits or transfers matter and information, encompassing techniques from conventional document printing to 3D volumetric displays.
• It integrates diverse modalities such as piezo-drive inkjet, fused deposition with syringe injection, electrohydrodynamic processes, and deterministic 2D stamping.
• Research advances optimize fidelity and control through high-resolution metrology, inverse nonlinear feedforward control, and innovative source attribution methods.

A printer, in the research literature surveyed here, denotes a device or system that places matter or encodes information onto a substrate, surface, or volume with controlled spatial registration. In the cited work, the term spans conventional document printers, piezo-drive inkjet platforms repurposed for full-colour volumetric displays, Fused Deposition Modeling systems augmented with syringe injection, electrohydrodynamic printers for large-area assembly of bio-piezoelectric films, computed axial lithography platforms, deterministic transfer tools for atomically layered materials, and mechatronic printer axes studied through inverse control and source-attribution methods (Hirayama et al., 2017, Khod et al., 2022, Joshi et al., 2021, Joshi et al., 2020, An et al., 21 Jul 2025, Waddell et al., 2 Sep 2025, Yan et al., 2023, Hemnani et al., 2018, Meer et al., 2021).

1. Printer as a research category

The cited literature treats printers as a heterogeneous but technically coherent class of systems whose common function is spatially selective placement. In document forensics, a printer is an electrophotographic source of intrinsic artifacts arising from toner spread, drum mechanics, and developer assembly; the central problem is attribution of a printed page to its source device (Joshi et al., 2020). In fabrication research, a printer is a deposition or transfer platform whose output may be a 3D fluorescent voxel stack, a magnetophoretic shell with fluid-filled cells, a bio-piezoelectric film, a volumetrically cured photopolymer part, or a deterministically placed 2D material flake (Hirayama et al., 2017, Yan et al., 2023, An et al., 21 Jul 2025, Waddell et al., 2 Sep 2025, Hemnani et al., 2018).

This breadth is reflected in the hardware. One study uses a commercially available piezo-drive inkjet printer (SO-KEN Inc., model TPW-105PB) “off the shelf,” with no custom print heads or mechanical parts added, to print photoreactive luminescence materials onto transparent films (Hirayama et al., 2017). Another uses a Creality Ender 3 Pro with a second “extruder” whose E-axis drives a 30 mL syringe through a NEMA 17 plus 8 mm lead screw, thereby combining polymer extrusion with fluid injection (Yan et al., 2023). OpenCAL uses an AAXA P6 Ultimate LED-based DLP projector, a 250 mm focal-length achromatic lens, a rotational stage, and a Raspberry Pi 5 to implement computed axial lithography (Waddell et al., 2 Sep 2025). MLSP uses conductive spiked metal disks mounted on an 18 G dispensing needle and driven at 7–10 kV to generate cone-jet emission (An et al., 21 Jul 2025). The “2D-Printer” uses a sharp micro-stamper and a transparent PDMS gel layer held by multi-axis micromanipulators for deterministic transfer (Hemnani et al., 2018).

A persistent misconception is that printers are only 2D output devices. The cited work directly contradicts this: the output may be a static 3D fluorescent image reconstructed from stacked slices, a volumetrically cured resin part, a large-area piezoelectric film assembled by aerosolized nanoscale droplets, or an arbitrary 3D object whose surface becomes an editable magnetophoretic display (Hirayama et al., 2017, An et al., 21 Jul 2025, Waddell et al., 2 Sep 2025, Yan et al., 2023).

2. Deposition, transfer, and volumetric encoding

The inkjet volumetric display of Hirayama et al. constructs a full-colour volumetric display from twenty transparent polyester films, each bearing a single slice of fluorescent voxels and stacked in register under 365 nm UV illumination (Hirayama et al., 2017). The red, green, and blue inks are a europium complex, a β-quinophthalone dye, and a coumarin dye with emission peaks of approximately 612 nm, 530 nm, and 470 nm, and quantum yields of approximately 0.43, 0.85, and 0.89, respectively. All three absorb in the near-UV at approximately 365 nm and re-emit in their characteristic primary colours. In the prototype, 300 × 300 px slices were printed over 35 × 35 mm, on 0.1 mm-thick polyester (Folex BG-32), with 0.5 mm inter-film gaps defined by acrylic spacers, giving a 35 × 35 × 12.5 mm stack. The prototype thus realizes up to 300 × 300 × 20 = 1.8 million physical voxels, while the native printer capability suggests an in-principle upper bound of 5,760 × 1,440 × 20 = 165 million. The associated multi-view design algorithm extends Nakayama et al. (2013) to full RGB by assigning each voxel a component-wise product of sampled colour values from multiple target projections, $$V_{(c)}(x,y,z) = A_{(c)}(u_a,v_a)\cdot B_{(c)}(u_b,v_b)\cdot C_{(c)}(u_c,v_c)$$, with output intensities obtained by summation along the corresponding projection path.

A different route to programmable printed appearance is the magnetophoretic display pipeline of Chan et al. A watertight mesh is converted into a thin shell filled with an array of self-contained cells, each produced by offsetting the mesh, remeshing the mid-surface, and lofting projected cross-sections between inner and outer shells (Yan et al., 2023). The base printer is a Creality Ender 3 Pro; the added injector uses a 14 Ga blunt needle with approximately 2.1 mm inner diameter, and Marlin 2.0.7 running on SKR 1.4 Turbo remaps the second E-axis to fluid injection. The injected mixture is empirically tuned to 25 wt% mineral oil, 35 wt% talcum powder, 40 wt% iron powder, and 1 wt% oil-based dye. Typical injection speed is approximately 10–20 mm$^3$/s, with retraction of approximately 1–2 mm of E1 and a purge dump volume of approximately 50 mm$^3$. Geometric constraints are explicit: $$d_{\mathrm{cell}}\in[2.5\,\mathrm{mm},6.5\,\mathrm{mm}]$$, $H_{\mathrm{cell}}\le 5$ mm, and outer screen thickness $H_{os}=0.6$–1 mm. In a 40 × 40 mm display, reducing $d_{\mathrm{cell}}$ from 6 mm to 3 mm increases cell count from approximately 49 to approximately 180.

Volumetric additive manufacturing appears in OpenCAL and MLSP, but with distinct physics. OpenCAL rotates a resin vial while projecting a sequence of 2D patterns; the cumulative dose is the angular backprojection of the projections, with local dose $D(r)=\int_0^T I(r,t)\,dt$, and curing when $D(r)\ge D_{\mathrm{th}}$ (Waddell et al., 2 Sep 2025). Its standard build volume is 30 mm diameter × 50 mm height, expandable to 100 mm diameter vials, with lateral resolution of approximately 40 $\mu$m, frame rate of approximately 15 Hz, rotation speed of approximately 3 RPM, and full-dose sequences of approximately 60 s. MLSP instead relies on electrohydrodynamic cone-jet generation: a DC high voltage of 7–10 kV applied across approximately 20 mm yields an electric field $^3$0; when electrical stress overcomes surface tension, a Taylor cone forms, and charged droplets undergo Coulomb fission after reaching the Rayleigh limit $^3$1 (An et al., 21 Jul 2025). With only two printheads at 8 kV and substrate heating, the measured throughput is approximately $^3$2, and the modular design makes the total rate scale approximately linearly with head count.

The “2D-Printer” is not a printer in the usual deposition sense but in the transfer sense. A sharp micro-stamper bends a 1–2 mm-thick PDMS layer locally and presses a selected flake into intimate, slow-peel contact with a target substrate (Hemnani et al., 2018). The substrate stage provides 1 $^3$3m or better X–Y resolution, and the achieved lateral placement accuracy is approximately $^3$4 (1$^3$5). When the flake is smaller than the stamper face, per-stamping yield is approximately 100%; when larger, only the region of direct contact transfers. This printer-like functionality is notable because it isolates placement from global substrate contamination.

3. Fidelity, metrology, and process optimization

The most explicit fidelity study in the cited corpus is the X-Ray CT characterization of four commercially available 3D printers: Ultimaker 2 Extended+, Delta Wasp, Raise E2, and ProJet MJP 3600 (Khod et al., 2022). Two CAD samples with designed porosity were printed: Sample 1, an 8 × 8 × 26 mm block with alternating cuboid cavities from 1.40 mm down to 0.20 mm and $^3$6; and Sample 2, a 6 × 6 × 17.5 mm block with alternating cubes and spheres and $^3$7. Thirty-eight printed replicas were scanned at 35 kV and 1 mA, reconstructed by Filtered Back Projection, and segmented with custom MATLAB codes. The principal metrics were porosity, $^3$8, porosity error, $^3$9, cusp density, $^3$0, and average surface roughness, $^3$1.

The resulting ranking is unambiguous. ProJet MJP in XHD mode prints voids down to 0.20 mm with approximately 100% fidelity and sharp boundaries, whereas the FDM systems reproducibly realize voids of at least 0.30 mm at approximately 50% size accuracy, with finer voids often distorted or absent (Khod et al., 2022). Measured porosity for Sample 1 / Sample 2 is approximately 35.2% / 27.5% for Ultimaker, 15.4% / 17.2% for Delta Wasp, 21.2% / 17.1% for Raise E2, and 14.6% / 14.6% for ProJet MJP. The minimum observed cusp-density proxy in XY slices is approximately $^3$2 for ProJet-XHD, compared with $^3$3 for Raise E2, $^3$4 for Delta Wasp, and $^3$5 for Ultimaker. The study therefore concludes that ProJet MJP gives the best quality of printed samples with the least amount of surface roughness and almost near to the actual porosity value, and that 100% infill density, minimum available layer height, and minimum nozzle speed give the best quality of 3D printing.

Other printer modalities expose different fidelity limits. In the inkjet volumetric display, film transmittance is approximately 82% at 365 nm and approximately 90% in the visible; after 20 clear films, UV excitation drops to approximately 2%, so brightness falls accordingly (Hirayama et al., 2017). Best image quality is at normal incidence; at $^3$6, blur appears because of parallax misalignment and inter-layer scattering, although the images remain clearly recognizable. OpenCAL reports lateral resolution of approximately 40 $^3$7m, vertical “voxel” resolution of approximately 200 $^3$8m equivalence, and radial runout below 100 $^3$9m (Waddell et al., 2 Sep 2025). MLSP reports droplet sizes reaching tens of nanometers and sub-$$d_{\mathrm{cell}}\in[2.5\,\mathrm{mm},6.5\,\mathrm{mm}]$$0m feature control, but also requires head-to-head spacing greater than approximately 7 cm to prevent mutual field suppression (An et al., 21 Jul 2025). These results suggest that fidelity in printer research is modality-specific: in some systems it is dominated by geometric void reproduction, in others by optical attenuation, electrostatic interference, or stage runout.

4. Printer dynamics and nonlinear feedforward control

A printer is also a controlled dynamical system. In the GP-based feedforward study, the plant is a consumer printer axis consisting of a current-driven DC motor, pulley, timing belt, and print-head carriage, sampled at 1 kHz and measured with a high-resolution incremental encoder (Meer et al., 2021). The nominal factory feedforward is linear, $$d_{\mathrm{cell}}\in[2.5\,\mathrm{mm},6.5\,\mathrm{mm}]$$1, and compensates only for nominal inertia and viscous drag. The core difficulty is nonlinear friction—static, Coulomb, and Stribeck—as well as position- and direction-dependent dead-zones induced by belt pre-tension.

The proposed solution models the inverse system as a noncausal nonlinear finite impulse response map, $$d_{\mathrm{cell}}\in[2.5\,\mathrm{mm},6.5\,\mathrm{mm}]$$2, with preview $$d_{\mathrm{cell}}\in[2.5\,\mathrm{mm},6.5\,\mathrm{mm}]$$3 samples and history $$d_{\mathrm{cell}}\in[2.5\,\mathrm{mm},6.5\,\mathrm{mm}]$$4 samples (Meer et al., 2021). A zero-mean Gaussian Process prior is placed on $$d_{\mathrm{cell}}\in[2.5\,\mathrm{mm},6.5\,\mathrm{mm}]$$5, typically using a stationary squared-exponential kernel $$d_{\mathrm{cell}}\in[2.5\,\mathrm{mm},6.5\,\mathrm{mm}]$$6, optionally augmented by a periodic kernel when friction repeats with belt pitch. Hyperparameters are identified by maximizing the marginal log-likelihood, and the predictive mean $$d_{\mathrm{cell}}\in[2.5\,\mathrm{mm},6.5\,\mathrm{mm}]$$7 generates the nonlinear feedforward signal for new tasks.

Experimentally, eleven scaled versions of a base third-order reference were run in closed loop, producing approximately 49,500 samples, then reduced to $$d_{\mathrm{cell}}\in[2.5\,\mathrm{mm},6.5\,\mathrm{mm}]$$8 training points by decimation (Meer et al., 2021). For the reference $$d_{\mathrm{cell}}\in[2.5\,\mathrm{mm},6.5\,\mathrm{mm}]$$9 seen in training, the linear feedforward yields $H_{\mathrm{cell}}\le 5$0 mm and $H_{\mathrm{cell}}\le 5$1 mm, while the GP-based feedforward yields 68 mm and 3.2 mm. For the unseen reference $H_{\mathrm{cell}}\le 5$2, the linear values are 164 mm and 4.2 mm, and the GP-based values are 86 mm and 3.6 mm. Maximum error near a friction “stiction” reversal around $H_{\mathrm{cell}}\le 5$3 s is reduced by a factor of 12. This establishes that printer performance cannot be reduced to printhead resolution or deposition physics alone; inverse modeling of carriage dynamics can be equally decisive.

5. Source printer identification and document forensics

Printer research also includes the inverse problem of identifying which printer produced a document. In the scanner-based study of Joshi et al., each connected component is represented by a printer-specific Local Texture Descriptor (PSLTD), a Local Binary Pattern-derived feature extracted from 3 × 3 neighborhoods using intensity similarity thresholds $H_{\mathrm{cell}}\le 5$4, gradient-direction threshold $H_{\mathrm{cell}}\le 5$5, pent-pattern binary vectors, and regrouped histograms over horizontal, vertical, forward-slant, and backward-slant structures (Joshi et al., 2021). The original representation is 10,502-dimensional; a reduced “approx. 4,602-dim” version retains identical cross-font performance at lower cost. To suppress intra-page variability, letter descriptors are pooled by columns or grids. In column pooling, the page interior is divided into $H_{\mathrm{cell}}\le 5$6 equal-width columns after margin trimming, and block descriptors are averaged as $H_{\mathrm{cell}}\le 5$7. Prediction can then be performed by block-wise Pearson correlation against printer references, followed by majority voting across blocks.

The reported cross-font results are strong but not uniform. On DB2, with training only on Cambria pages, proposed grid pooling (8 × 8) with correlation achieves 99.5% for Cambria-to-Cambria, 93.5% for Cambria-to-Arial, 94.3% for Cambria-to-Times New Roman, and 60.3% for Cambria-to-Comic Sans (Joshi et al., 2021). For the full 18-printer set, same-font accuracy reaches 100% for both pooling variants; intra-model identification of three identical Canon MF-4820d printers is also 100% correct. The method is intrinsic and passive: no watermark or extrinsic tag is embedded. Its main limitations are dependence on high-precision 16-bit scans for the correlation-based variant, degradation on radically different fonts, and the need for approximately 15–20 letters per block.

The smartphone-based study extends printer attribution to distributed acquisition (Joshi et al., 2020). It introduces a dataset of 2250 document images acquired from eighteen printers, each contributing twenty-five Cambria pages captured at five settings: 0° tilt, +5° tilt, –5° tilt, free-hand, and low illumination. Letter patches are modeled as $H_{\mathrm{cell}}\le 5$8, with an ideal image $H_{\mathrm{cell}}\le 5$9 estimated from histogram-based partitions $H_{os}=0.6$0, and residual $H_{os}=0.6$1. A two-channel CNN takes the native letter patch and its residual as $H_{os}=0.6$2 input, using two 3 × 3 convolutional layers with BatchNorm and ReLU, max-pooling, a 512-unit dense layer, and an 18-way softmax. With Adam at learning rate 0.001, weight decay 0.0005, batch size 100, and 50 epochs, the system achieves 98.42% document classification accuracy using images of the letter “e” under a 5 × 2 cross-validation approach. When tested on approximately half a million letters of all types, it reaches 90.33% letter-level and 98.01% document-level accuracy. Performance across acquisition settings remains high—98.12% at +5°, 97.86% free-hand, 97.69% in low illumination—but falls to 94.70% at –5°, partly because OCR fails under strong perspective distortion.

Taken together, these studies establish two important points. First, printer fingerprints remain detectable after scanning and even after smartphone acquisition. Second, attribution quality depends strongly on representation and acquisition protocol: pooled handcrafted descriptors with correlation excel in controlled 16-bit settings, whereas learned fusion of letter appearance and residuals is robust under smartphone imaging (Joshi et al., 2021, Joshi et al., 2020).

6. Applications, constraints, and research implications

The application space of printer research in these papers is unusually broad. The inkjet volumetric display is proposed for digital signage, media art, entertainment, and security, and demonstrates full-colour 3D “butterfly” and “flower” objects as well as multi-view displays projecting three independent photos or four patterns (Hirayama et al., 2017). MLSP targets large-area β-glycine films for biocompatible sensors, energy harvesters, actuators, and wearable devices, with effective out-of-plane $H_{os}=0.6$3 pm/V by PFM measurement and no need for post-processing polarization (An et al., 21 Jul 2025). OpenCAL is designed for academic makerspaces, hands-on education in emerging manufacturing technologies, photopolymer science, and computational imaging, with documented CAD, firmware, and user materials (Waddell et al., 2 Sep 2025). The magnetophoretic display pipeline enables a Stanford bunny with customizable appearances, an espresso mug usable as a post-it note surface, a board game figurine with a computationally updated display, and flexible wearable accessories with editable visuals (Yan et al., 2023). The “2D-Printer” is tested on Silicon photonic chips and is proposed as a route to rapid device prototyping because of its contamination-free process and substrate benignity (Hemnani et al., 2018). In forensics, source printer identification is directly linked to forged contracts, anonymous letters, financial statements, chain of custody, and remote distributed analysis (Joshi et al., 2020).

The constraints are equally specific. The layered fluorescent display suffers exponential attenuation of UV excitation and visible emission as film count increases, and its current depth resolution is limited by 0.5 mm interlayer spacing (Hirayama et al., 2017). MLSP requires high-voltage power at 7–10 kV, careful insulation, controlled viscosity and conductivity in a 10:1:2.5 water:glycine:ethanol formulation, roll-to-roll heating, and sufficient spacing between heads (An et al., 21 Jul 2025). OpenCAL requires post-processing for uncured resin removal; its solvent-based cleaning and centrifugal cleaning modules exist precisely because volumetric curing does not eliminate downstream handling problems (Waddell et al., 2 Sep 2025). The magnetophoretic display currently offers only two colours and depends on cell-opening and overhang constraints during FDM fabrication (Yan et al., 2023). The 2D-material printer is still manually operated at 1–5 minutes per flake, although the stamping action itself is less than 1 s (Hemnani et al., 2018). The GP feedforward method has cubic training complexity in the number of retained samples and therefore motivates sparse or local GP variants for larger datasets (Meer et al., 2021). Forensics systems remain sensitive to font shift, OCR failure, and acquisition precision (Joshi et al., 2021, Joshi et al., 2020).

A plausible implication is that printer research is converging toward integrated stacks in which material formulation, optics, mechanics, signal processing, and inference are inseparable. The cited work does not describe a single canonical printer architecture; instead, it shows that “printer” has become a systems concept encompassing deposition physics, deterministic placement, inverse metrology, nonlinear control, and forensic attribution.