# 3D printing > Source: https://aiwiki.ai/wiki/3d_printing > Updated: 2026-06-24 > Categories: Generative AI, Robotics > From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution. **3D printing**, also called **additive manufacturing** (AM), is a family of fabrication processes that build three-dimensional objects from a digital model by adding material in successive layers, the opposite of subtractive methods like milling that cut material away. The term covers seven distinct process categories standardized by ISO/ASTM 52900, ranging from photopolymer resins cured by ultraviolet light to metal powders melted by lasers and electron beams. Materials handled commercially include thermoplastics, photopolymers, metal alloys, ceramics, sand, concrete, food, and living cells. The global market for additive manufacturing products and services reached roughly 21.9 billion U.S. dollars in 2024, up 9.1 percent year over year, and roughly 24.2 billion dollars by the 2026 reporting cycle.[1][2] The technology started in the mid 1980s as a way to make engineering prototypes faster than machining could. It stayed niche through the 1990s. Two things changed it. The expiry of foundational fused-deposition patents in the late 2000s opened up cheap desktop machines, and the maturing of metal powder bed fusion in the 2010s let aerospace and medical companies print parts that could not be made any other way. By the early 2020s, additive manufacturing had become a standard tool in aerospace, dental, hearing aids, orthopedic implants, jewelry, footwear, and rapid prototyping. From an AI perspective, 3D printing matters for three reasons. First, it is one of the largest commercial applications of [generative design](/wiki/generative_design) and topology optimization, both of which depend on numerical optimization and increasingly on [machine learning](/wiki/machine_learning) surrogates. Second, modern metal printers ship with [computer vision](/wiki/computer_vision) systems that monitor the melt pool layer by layer to catch defects in process, a closed-loop inspection problem now dominated by [convolutional neural networks](/wiki/convolutional_neural_network). Third, a wave of generative AI tools that turn text prompts or 2D images into printable 3D meshes (OpenAI's [Point-E](/wiki/point_e) and [Shap-E](/wiki/shap_e), Tencent's [Hunyuan3D](/wiki/hunyuan_3d), [Tripo](/wiki/tripo), Meshy, and Luma Genie among them) connects 3D printing to the same generative stack that produces images and video. This makes 3D printing a recurring case study in [AI for science](/wiki/ai_for_science) and [AI in manufacturing](/wiki/ai_in_manufacturing). ## History ### When was 3D printing invented? The modern lineage begins in 1984. In May of that year, three French inventors (Alain Le Mehaute, Olivier de Witte, and Jean-Claude Andre) filed a patent for stereolithography that their employers, CILAS and CGE, chose not to pursue. Three weeks later, in August 1984, Charles Hull filed his stereolithography patent in California. Hull's process used a vat of liquid photopolymer, an ultraviolet laser tracing each cross-section, and a build platform that stepped down between layers. The patent was granted in March 1986.[3] That same year Hull co-founded 3D Systems Corporation in Valencia, California, which shipped the SLA-1 in 1987. The .stl file format, still the dominant input format for 3D printers four decades later, comes from this machine. In the mid 1980s, Carl Deckard, an undergraduate at the University of Texas at Austin, developed selective laser sintering (SLS) under the supervision of Joe Beaman. Their patent was filed in 1986 and granted in 1989.[4] Deckard and Beaman founded what became DTM Corporation, which 3D Systems acquired in 2001. In 1988, Scott Crump invented fused deposition modeling (FDM), a process that extrudes molten thermoplastic through a heated nozzle. He and his wife Lisa Crump founded Stratasys in Eden Prairie, Minnesota in 1989. The FDM patent (US 5,121,329) issued in 1992.[5] The expiry of this patent in 2009 unlocked the desktop 3D printing market. ### The desktop wave In 2005 Adrian Bowyer, a senior lecturer in mechanical engineering at the University of Bath, started the RepRap Project, an open-source effort to design a 3D printer that could print most of its own parts. The first working RepRap, *Darwin*, was completed in 2007, and RepRap derivatives became the basis for almost every consumer FDM machine that followed.[6] MakerBot Industries was founded in Brooklyn in 2009 by Bre Pettis, Adam Mayer, and Zach Smith. The Cupcake CNC kit shipped that year for around 750 U.S. dollars; Stratasys acquired MakerBot in 2013. Josef Prusa, a contributor to the RepRap project since 2009, launched what became Prusa Research in Prague; the Prusa i3 design, released in 2012, became the most-cloned FDM printer in history. Formlabs was founded in 2011 in Somerville, Massachusetts by three MIT Media Lab graduates: Maxim Lobovsky, Natan Linder, and David Cranor. Their Form 1 desktop SLA printer raised about 2.95 million U.S. dollars on Kickstarter in 2012. Carbon (originally Carbon3D) was founded in 2013 in Redwood City, California by Joseph and Philip DeSimone, Edward Samulski, and Alex Ermoshkin. In March 2015, Joseph DeSimone unveiled CLIP (Continuous Liquid Interface Production) at TED. The process uses an oxygen-permeable window to suppress curing in a thin dead zone above it, allowing the print to be drawn out of the resin continuously rather than layer by layer.[7] The commercial system, Digital Light Synthesis (DLS), shipped in 2016. Carbon partnered with Adidas to produce the Futurecraft 4D midsole, announced in 2017. Desktop Metal was founded in 2015 in Burlington, Massachusetts by Ric Fulop and several MIT faculty; it went public via SPAC in 2020. Bambu Lab, founded in Shenzhen in 2020 by former DJI engineers, shipped the X1 Carbon in 2022. The machine prints at toolhead speeds up to 500 millimeters per second with active vibration compensation and pulled the desktop FDM market into a generation that had stayed largely Prusa-shaped since the early 2010s. ## What are the seven process families (ISO/ASTM 52900)? The joint ISO/ASTM 52900 standard, first issued in 2015 and updated as 52900:2021, defines seven process categories.[8] Every commercial 3D printer falls into one of these. | Process category | Common names | Material | Heat source / mechanism | |---|---|---|---| | Vat photopolymerization | SLA, DLP, LCD/MSLA, CLIP, DLS | Photopolymer resin | UV laser (SLA), DLP projector, LCD mask, oxygen-permeable window (DLS) | | Material extrusion | FDM, FFF | Thermoplastic filament, pellets, paste | Heated nozzle | | Powder bed fusion | SLS, MJF, LPBF/SLM, DMLS, EBM | Polymer powder, metal powder | Laser (SLS, LPBF), IR lamp + agent (MJF), electron beam (EBM) | | Binder jetting | BJ | Sand, metal powder, ceramic, gypsum | Inkjet binder + later sintering or infiltration | | Material jetting | PolyJet, MultiJet | Photopolymer droplets, wax | UV cure on jetted droplets | | Directed energy deposition | DED, LENS, EBAM, WAAM | Metal powder or wire | Focused laser, electron beam, or arc | | Sheet lamination | LOM, UAM | Paper, plastic, metal foil | Adhesive, ultrasonic welding | Vat photopolymerization gives the highest surface finish and dominates dental, jewelry, and hearing-aid production. Material extrusion is cheapest and most common for prototyping and consumer use. Powder bed fusion, especially laser powder bed fusion (LPBF, also called SLM or DMLS depending on the vendor), is the workhorse for metal end-use parts in aerospace and medical. Binder jetting is faster per part because it does not melt anything during the print, instead gluing powder together for later sintering. Material jetting (Stratasys PolyJet) is used for full-color and multi-material prototypes. Directed energy deposition adds material to existing parts and builds large structures; Wire Arc Additive Manufacturing has been used for ship propellers and pedestrian bridges. Sheet lamination is rare in modern production. ## Materials ### Thermoplastics FDM and SLS together cover most of the polymer market. Common thermoplastics include polylactic acid (PLA), the easiest to print but biodegradable and with a low glass transition temperature; acrylonitrile butadiene styrene (ABS), tougher and heat-resistant but prone to warping; polyethylene terephthalate glycol (PETG), a middle-ground material; nylon (polyamide 6, 11, and 12), the standard SLS material; and engineering thermoplastics like polycarbonate, polyetheretherketone (PEEK), and polyetherketoneketone (PEKK) used in aerospace interior parts and medical implants. PEEK in particular is FDA-cleared for cranial and spinal implants. ### Photopolymers Vat photopolymerization processes use mixtures of monomers, oligomers, and photoinitiators that cross-link under UV light. Resin chemistry has expanded from brittle prototyping resins to engineering grades approximating ABS, polypropylene, and rubber, plus medical-grade resins for surgical guides and dental aligners. Carbon's DLS materials, which include EPU 41, RPU 70, and EPX 82, achieve higher mechanical properties than typical SLA resins by combining photo-curing with a secondary thermal cure. ### Metals The metal additive market is concentrated around a small set of alloys: titanium alloy Ti-6Al-4V (used for aerospace structures and medical implants), nickel superalloys Inconel 625 and 718 (turbine and combustion-chamber parts), aluminum alloy AlSi10Mg (lightweight aerospace and automotive parts), 17-4 PH and 316L stainless steels, cobalt-chrome (dental and orthopedic), and tool steels like H13 and maraging steel. Powder for these alloys is typically gas-atomized and sieved to 15 to 53 microns for laser powder bed fusion or 45 to 105 microns for electron beam melting. ### Ceramics, concrete, food, and bio-inks Ceramic 3D printing covers technical ceramics (alumina, zirconia, silicon carbide) and traditional materials like porcelain. Concrete printing is used in construction (ICON, Apis Cor, COBOD). Bio-inks, mixtures of cells, hydrogels, and growth factors, are extruded by bioprinters; the field overlaps with tissue engineering and is mostly research-stage. ## Major manufacturers | Company | Founded | Headquarters | Flagship technology | |---|---|---|---| | 3D Systems | 1986 | Rock Hill, South Carolina, USA | SLA, SLS, MJP, DMP | | Stratasys | 1989 | Eden Prairie, Minnesota / Rehovot, Israel | FDM, PolyJet | | EOS | 1989 | Krailling, Germany | LPBF, polymer SLS | | 3D Systems / DTM | acquired 2001 | (legacy SLS line) | SLS | | Renishaw | 1973 (founded), AM since 2011 | Wotton-under-Edge, UK | LPBF | | SLM Solutions (now part of Nikon) | 2007 | Lubeck, Germany | Multi-laser LPBF | | Formlabs | 2011 | Somerville, Massachusetts, USA | Desktop SLA, SLS | | Carbon | 2013 | Redwood City, California, USA | DLS (vat photopolymerization) | | Markforged | 2013 | Waltham, Massachusetts, USA | FDM with continuous-fiber reinforcement, metal extrusion | | Desktop Metal | 2015 | Burlington, Massachusetts, USA | Bound metal deposition, binder jetting | | Velo3D | 2014 | Fremont, California, USA | Support-free LPBF | | HP Inc. (3D printing) | 2014 (entered AM) | Palo Alto, California, USA | Multi Jet Fusion (polymer), Metal Jet (binder jetting) | | Prusa Research | 2012 | Prague, Czech Republic | Open-source FDM, Original Prusa series | | Bambu Lab | 2020 | Shenzhen, China | High-speed FDM with input shaping | | Anycubic / Elegoo / Creality | various, 2010s | China | Low-cost FDM and MSLA | GE Aerospace is not a printer manufacturer in the consumer sense, but it is one of the largest single buyers of metal AM machines. Its LEAP engine fuel nozzle, certified by the FAA in 2015 and produced via direct metal laser melting in Inconel 718, consolidated 20 parts into one, is 25 percent lighter, and lasts up to five times longer than the brazed assembly it replaced; it is generally cited as the first widely produced AM part in commercial aviation.[9] GE passed 100,000 printed LEAP nozzles by 2020.[9] ## How is AI used in 3D printing? AI shows up across the 3D printing pipeline. Some of it is genuine machine learning. Some of it is older numerical optimization rebranded. The line between the two is blurry on purpose because the marketing pressure to call everything AI has been unusually strong in this industry. The clearest, best-documented intersections are four: generative design and topology optimization, in-situ defect detection by computer vision, machine learning for process and materials qualification, and generative text-to-3D model creation. ### What is generative design in 3D printing? [Topology optimization](/wiki/topology_optimization) is the older idea: given a design space, a load case, and a set of constraints, find the distribution of material that minimizes compliance (or some other objective) subject to a volume fraction. The method, formalized by Bendsoe and Kikuchi in 1988, predates the AI rebrand by a long way.[10] Modern implementations use SIMP (Solid Isotropic Material with Penalization) or level-set methods. [Generative design](/wiki/generative_design), as marketed by Autodesk and others since around 2015, is a related but broader idea: explore a population of design candidates that satisfy the same constraints, often using shape grammars, evolutionary algorithms, or topology optimization with multiple starting conditions. The user gets a set of candidates rather than a single optimum and picks among them. - Autodesk Fusion 360 Generative Design (formerly Project Dreamcatcher, integrated 2018) runs on Autodesk's cloud and can target multiple manufacturing processes (additive, milling, casting) within the same problem. - nTop (formerly nTopology), founded in 2015, is built on implicit modeling (signed distance functions and lattice structures) rather than B-rep geometry, which lets it represent complex internal structures without exploding the file size. - Altair Inspire and OptiStruct are the long-running topology optimization tools from Altair, used heavily in automotive structures. - Siemens NX includes generative engineering modules, and PTC Creo integrated Frustum's solver after acquiring Frustum in 2018. - Hyperganic (Munich, founded 2015) and ParaMatters (now part of CASTOR) push algorithmically generated geometry that is hard to design by hand. The outputs of these tools, organic-looking lattice structures and bone-like trusses, are difficult or impossible to manufacture by milling, which is why they show up next to additive manufacturing so often. Bracket consolidation is the canonical case study: a widely cited 2018 General Motors and Autodesk demonstration produced a seat-bracket design that was 40 percent lighter and 20 percent stronger than the original and consolidated eight stamped-and-welded components into a single printed part, with the software generating more than 150 valid candidate designs from the engineers' constraints.[11] GE, Airbus, and others have shown comparable 30 to 50 percent mass reductions on individual brackets with topology optimization plus LPBF. Machine learning enters this stack in two main ways. First, neural networks are used as surrogates to replace expensive finite-element evaluations during optimization. Papers in *Additive Manufacturing*, *Computer-Aided Design*, and *Structural and Multidisciplinary Optimization* have demonstrated [graph neural networks](/wiki/graph_neural_network) and convolutional networks that approximate stress fields several orders of magnitude faster than direct FEM. Second, generative models ([variational autoencoders](/wiki/variational_autoencoder), [GANs](/wiki/generative_adversarial_network), [diffusion models](/wiki/diffusion_model)) are used to propose new geometry families, with the topology optimizer as the loop's reward function. ### How does AI detect 3D printing defects in real time? Laser powder bed fusion is fundamentally a melt-pool process: a laser sweeps across a powder bed at hundreds of millimeters per second, melting tracks roughly 100 microns wide. If the laser power, scan speed, layer thickness, gas flow, or powder spreadability is off, the melt pool destabilizes and the resulting porosity, lack of fusion, or keyholing all show up as defects in the finished part. For aerospace and medical applications where every part needs traceability, that is a problem. Most industrial LPBF machines now ship with some form of in-situ monitoring: - **Photodiode and pyrometer melt-pool monitoring**: dichroic mirrors send a fraction of the laser-induced emission to photodiodes, which produce a time-series signal correlated with melt-pool size. EOS sells this as EOSTATE Meltpool, SLM Solutions as MPM (Melt Pool Monitoring), and Sigma Additive Solutions (formerly Sigma Labs) markets PrintRite3D for retrofit. - **High-speed imaging**: coaxial or off-axis cameras image the melt pool at kilohertz frame rates. Vendor systems include EOS's coaxial camera and Concept Laser's QM Meltpool. - **Layer-wise imaging**: low-resolution cameras shoot the powder bed after each recoat, looking for short-feeds, recoater streaks, and recoater interruptions. Most machines now do this. - **Acoustic emission**: piezoelectric sensors on the build plate pick up high-frequency emissions correlated with cracking and keyholing. AI enters the loop because the raw signals from these sensors are noisy and high-bandwidth, reaching hundreds of gigabytes per build for a multi-laser machine. [Convolutional neural networks](/wiki/convolutional_neural_network) and recurrent networks have been applied to classify melt-pool images and time-series into stable, lack-of-fusion, and keyhole regimes. Peer-reviewed studies report defect-detection accuracy above 93 percent from in-situ layer-wise images and multi-spectral emission, and a CNN verified against X-ray computed tomography reached 94 percent accuracy across six classes of unseen defective regions.[12] Supervised models trained on intentionally seeded defects can flag suspicious regions for X-ray CT inspection after the build, which is the only widely accepted method for porosity verification but is slow and expensive. Closed-loop control, where the machine adjusts laser power based on the melt-pool signal in real time, exists in research demonstrations and limited commercial form (Velo3D's Sapphire systems, EOS's Smart Fusion, and Inkbit's Vista jetting platform are among the most cited). For large-format extrusion and DED, where layers take seconds rather than milliseconds, [computer vision](/wiki/computer_vision) is mostly used for shape conformance: stereo or structured-light scanning between layers, comparing the as-built point cloud to the design, flagging deviations. ### Machine learning for materials and process parameters LPBF qualification is expensive: a single new alloy might take a year of design-of-experiments work to map process windows. [Bayesian optimization](/wiki/bayesian_optimization), Gaussian-process-based active learning, and gradient-boosted regression on previous experiments have shrunk this. Papers from Argonne, NIST, and the Singapore Centre for 3D Printing have shown reductions of 50 to 80 percent in the number of experiments needed to find a process window for a new alloy. ML is also used in: - **Powder QC**: convolutional networks classify scanning electron microscope images of powder for sphericity, satellite content, and contamination. - **Slicer optimization**: ML-based slicer plugins in Cura and PrusaSlicer attempt to predict warping and adjust orientation or support placement automatically. - **Post-processing**: deep learning surrogates for thermal post-processing simulate hot isostatic pressing (HIP) and heat treatment cycles to predict residual stress and microstructure. Open datasets exist for some of this work: the AM-Bench data from NIST and the Penn State / GE TRIP-PBF dataset are the most widely used. ### Can AI generate 3D models from text? (text-to-3D and generative geometry) A newer category of tools attempts to turn natural-language prompts into editable CAD or printable mesh geometry. This is the area where the AI hype most closely matches reality, since the underlying models are large language models and [diffusion models](/wiki/diffusion_model) trained on geometry datasets. - **Zoo (zoo.dev)**, founded 2020 in San Francisco, has released *Text-to-CAD*, a transformer-based model that emits parametric KCL (their own scripting language) from natural-language prompts. It supports basic part shapes and features rather than full assemblies. - **Spline AI** offers text-to-3D inside its browser-based 3D editor, mostly aimed at design and motion graphics rather than engineering. - **CADCrafter, DeepCAD, and SkexGen** are research projects that train transformers on B-rep CAD sequences, treating the construction history as a token sequence. They can generate novel parts within their training distribution but are not yet at engineering quality. For mesh generation rather than parametric CAD, the field has moved fast since 2022. The table below summarizes the most-cited generative 3D systems and their verified release details. | Model | Origin | Released | What it does | |---|---|---|---| | [Point-E](/wiki/point_e) | OpenAI | Dec 2022 | Text-to-3D point clouds; first a synthetic image via GLIDE, then a point-cloud diffusion stack; about 1 to 2 minutes on a single GPU | | [Shap-E](/wiki/shap_e) | OpenAI | May 2023 | Generates implicit functions renderable as textured meshes or [NeRFs](/wiki/nerf); about 13 seconds on a single NVIDIA V100 | | [Hunyuan3D-2](/wiki/hunyuan_3d) | Tencent | Jan 2025 | Open-source two-stage system: Hunyuan3D-DiT geometry (2.6B parameters) plus Hunyuan3D-Paint texture (1.3B parameters) | | [Tripo](/wiki/tripo) | VAST (Beijing) | 2023 onward | Text- and image-to-3D for game and film assets; parent VAST passed a 1 billion dollar valuation after a March 2026 round | | Meshy | Meshy Inc. | 2023 onward | Textured meshes from text or a single image via multi-stage diffusion plus retopology | | Luma AI Genie | Luma AI | 2023 | Meshes from text prompts | | Stable 3D / Stable Fast 3D | Stability AI | 2023 to 2024 | Image-to-3D from a single photograph | | InstantMesh, TRELLIS, CRM | Research labs | 2024 | Generate or sparse-view-reconstruct meshes in seconds | OpenAI's Point-E, released in December 2022, was framed explicitly as a speed-versus-quality trade: "While our method still falls short of the state-of-the-art in terms of sample quality, it is one to two orders of magnitude faster to sample from."[13] Shap-E followed in May 2023 and, unlike Point-E, "directly generates the parameters of implicit functions that can be rendered as both textured meshes and neural radiance fields."[14] Tencent's Hunyuan3D-2, released open-source in January 2025, pairs a 2.6-billion-parameter geometry diffusion transformer (Hunyuan3D-DiT) with a 1.3-billion-parameter texture model (Hunyuan3D-Paint).[15] Meshy reported reaching 15 million dollars in annual recurring revenue, and Tripo's Beijing-based parent VAST crossed a 1 billion dollar valuation after a 50 million dollar round led by Alibaba in March 2026.[16][17] For 3D printing specifically, the catch is that decorative meshes from these tools are usually not directly printable. They tend to have non-manifold geometry, internal cavities, and flying islands that violate slicer assumptions. A practical pipeline runs the generated mesh through a repair step (Meshmixer, Microsoft 3D Builder, Materialise Magics) before slicing. Hobbyists print [Midjourney](/wiki/midjourney)- or Meshy-generated figurines this way; engineering use is rare so far because the geometry is not parametric and cannot be edited cleanly. ### Robotics and prosthetics 3D printing and [robotics](/wiki/robotics) interact heavily, especially for soft robotics and prosthetics. Soft robots, which use deformable elastomers and pneumatic chambers rather than rigid links, are difficult to manufacture by molding because the internal chambers and channels twist around each other in ways that demolding cannot handle. Researchers at Harvard's Wyss Institute, MIT CSAIL, and ETH Zurich have used multi-material jetting (Stratasys J-series and Inkbit) to print soft pneumatic actuators in a single build with both rigid frames and soft bellows. The Octobot, published in *Nature* in 2016, was an early demonstration of an entirely soft, 3D-printed autonomous robot; the Harvard team described it as "the first autonomous, untethered, entirely soft robot," with movement driven pneumatically by gas from a hydrogen-peroxide reaction and routed by a printed microfluidic logic circuit.[18] Prosthetic and orthotic work runs on a more direct economic logic. Custom-fitted prosthetic sockets, traditionally made by plaster casting and lamination, can be 3D-scanned and FDM-printed in days rather than weeks. The non-profit e-NABLE community has distributed open-source 3D-printable hand prostheses since 2013, mostly for children, who outgrow their devices quickly. Commercial prosthetics from companies like Mecuris and Unlimited Tomorrow use the same workflow with higher-grade materials. For robot bodies more broadly, FDM and SLS parts are common in research robots and in low-volume commercial designs. [Boston Dynamics](/wiki/boston_dynamics)' [Spot](/wiki/spot_robot) uses several SLS nylon parts. Open-source quadrupeds (Stanford Doggo, MIT Mini Cheetah derivatives) and humanoid hands rely heavily on printed parts because the iteration speed is too high for machined components. ### Bio-printing Bio-printing extrudes or jets bio-inks (cells in a hydrogel carrier, with growth factors) to build tissue scaffolds. The field is mostly research as of 2025, with no fully functional printed organ in clinical use. Major players include Organovo, CELLINK (now part of BICO), Allevi, Aspect Biosystems, and Prellis Biologics. Printed skin grafts and cartilage have reached limited clinical trials. Vascularization, the problem of building blood vessels small enough and dense enough to keep printed tissue alive, remains the bottleneck. ## Industry applications ### Aerospace Aerospace was the first heavy industry to certify printed metal parts for production aircraft. The GE Aviation LEAP fuel nozzle (entered service 2016 on the Airbus A320neo and Boeing 737 MAX) consolidated 20 brazed parts into one printed Inconel piece, weighing 25 percent less and lasting roughly five times longer.[9] Pratt & Whitney, Rolls-Royce, Honeywell, and Safran all run production AM lines for combustor and turbine components. SpaceX prints chambers and injectors for SuperDraco and Raptor engines. Relativity Space, founded in 2015, has built some of the world's largest metal 3D printers (the Stargate series) for printing entire rocket structures; its Terran 1 launch vehicle, roughly 85 percent printed by mass, lifted off on March 22, 2023 and reached space but fell short of orbit, the first largely 3D-printed rocket to fly.[19] ### Medical and dental Dental aligners are probably the highest-volume application of 3D printing. Align Technology (Invisalign), founded 1997, prints tens of millions of aligner molds per year on banks of Stratasys and 3D Systems machines, then thermoforms the aligners over the printed molds. Hearing aid shells are printed by Sonova, Starkey, and others; the audiology industry switched almost entirely from manual fabrication to SLA between 2003 and 2008. In orthopedics, titanium acetabular cups, spinal cages, and cranial plates have FDA clearance from companies including 4WEB Medical, Renovis, Stryker, and Smith+Nephew. The lattice surfaces on these implants encourage bony in-growth in a way that machined surfaces cannot. PEEK 3D-printed cranial plates, available from companies like Restor3d, are used for skull reconstructions. Surgical guides, patient-specific anatomical models for surgical planning, and even single-use surgical instruments are printed routinely now. The Mayo Clinic, Cleveland Clinic, and Royal National Orthopaedic Hospital have in-house point-of-care 3D printing labs. ### Footwear and consumer goods Adidas Futurecraft 4D, announced in 2017 with Carbon, was the first widely sold 3D-printed midsole. Subsequent Carbon collaborations with New Balance, Riddell (helmet liners), and Specialized (cycling saddles) made digital lattice midsoles common in performance footwear. HP's Multi Jet Fusion has printed eyewear, watch components, and consumer electronics housings. ### Construction Concrete 3D printing has moved from one-off demonstrations to small commercial deployment. ICON, founded 2017 in Austin, Texas, has printed a small number of houses in collaboration with Lennar and 3Strands. Apis Cor (Russia/U.S.) and COBOD (Denmark) make portal-style concrete printers. WinSun in China has printed apartment buildings and offices. The economics are still marginal compared to conventional construction in most markets; the technology's strongest case is in disaster relief and remote builds. ### Automotive Automotive use is mostly tooling, jigs, and fixtures rather than production parts. BMW, Bugatti, and Porsche have printed limited-run end-use parts (notably the Bugatti Chiron's titanium brake calipers). Spare parts on demand for older vehicles is a recurring promise that has not yet broken through to volume. ## Limitations 3D printing is not a replacement for injection molding at high volumes. The break-even point depends on the part, but for commodity polymer parts under 100 grams, injection molding wins on cost above a few thousand units. Build rates for metal LPBF top out at roughly 50 to 200 cubic centimeters per hour per laser; multi-laser machines from EOS, SLM, Velo3D, and others bring this up but not to the levels of casting or forging. Surface finish from most processes is poor without post-processing. As-printed FDM has visible layer lines. SLS has a slight powdery texture. LPBF metal parts have surface roughness Ra in the range of 6 to 25 micrometers and almost always need machining on critical surfaces. Residual stress in metal AM is significant. Parts often need stress-relief heat treatment before they come off the build plate, plus support-removal machining and frequently hot isostatic pressing. Material selection is narrower than in conventional manufacturing. There is no 3D-printable equivalent of cast iron, mass-produced thermoset rubber, or cheap commodity steel. Anisotropy, the difference in mechanical properties along versus across build direction, is a persistent problem for FDM and a smaller but real one for LPBF. Part traceability and qualification are heavy. The aerospace AM standard AS9100 and ASTM F3303 set the bar for documentation. A single LPBF aerospace part can have hundreds of pages of process records. The AI defect-detection and qualification methods above are partly an answer to this cost: faster process-window mapping and in-situ inspection are how the industry hopes to cut the per-part paperwork burden. ## Standards and quality The core standards are maintained jointly by ISO Technical Committee 261 and ASTM Committee F42. Key documents: - **ISO/ASTM 52900:2021**: terminology and process categories. - **ISO/ASTM 52901:2017**: requirements for purchased AM parts. - **ISO/ASTM 52902:2023**: standard test artifacts for benchmarking. - **ISO/ASTM 52904:2024**: practice for laser powder bed fusion process qualification. - **ISO/ASTM 52911 series**: design for additive manufacturing. - **ASTM F3055, F3056, F3091, F3184, F3187, F3213, F3301, F3302, F3303**: materials and process specifications for specific alloys (Inconel 625/718, Ti-6Al-4V, 316L, AlSi10Mg) and DED. - **AS9100D / AS9100 Rev D**: aerospace quality management system, applied to AM facilities. - **FDA guidance on 3D printed medical devices** (2017, updated 2023): the technical considerations document covers design, software, post-processing, and validation for printed medical devices.[20] The Wohlers Report, published annually since 1996 by Wohlers Associates (part of ASTM International since 2021), is the standard market and technology reference for the industry. The Wohlers Report 2025 put the 2024 market at about 21.9 billion dollars (9.1 percent growth), and the 2026 edition reported revenues of about 24.2 billion dollars.[1][2] ## See also - [Generative design](/wiki/generative_design) - [Topology optimization](/wiki/topology_optimization) - [Computer vision](/wiki/computer_vision) - [Convolutional Neural Network](/wiki/convolutional_neural_network) - [Robotics](/wiki/robotics) - [Machine learning](/wiki/machine_learning) - [AI in manufacturing](/wiki/ai_in_manufacturing) - [AI for science](/wiki/ai_for_science) - [Point-E](/wiki/point_e) - [Shap-E](/wiki/shap_e) - [Hunyuan 3D](/wiki/hunyuan_3d) - [Tripo](/wiki/tripo) - [NeRF](/wiki/nerf) - [Diffusion model](/wiki/diffusion_model) ## References 1. Wohlers Associates / ASTM International. *Wohlers Report 2025: 3D Printing and Additive Manufacturing State of the Industry*, 2025 (reports a 21.9 billion U.S. dollar market for 2024, up 9.1 percent). https://wohlersassociates.com 2. TCT Magazine. "Wohlers Report 2026: Additive manufacturing revenues reach $24.2 billion," 2026. https://www.tctmagazine.com/wohlers-report-2026-additive-manufacturing-revenues-reach-24-2-billion/ 3. Hull, Charles W. "Apparatus for Production of Three-Dimensional Objects by Stereolithography." U.S. Patent 4,575,330, filed August 8, 1984, granted March 11, 1986. 4. Deckard, Carl R. "Method and Apparatus for Producing Parts by Selective Sintering." U.S. Patent 4,863,538, filed October 17, 1986, granted September 5, 1989. 5. Crump, S. Scott. "Apparatus and Method for Creating Three-Dimensional Objects." U.S. Patent 5,121,329, filed October 30, 1989, granted June 9, 1992. 6. RepRap Project documentation. https://reprap.org/wiki/Main_Page 7. Tumbleston, J.R. et al. "Continuous Liquid Interface Production of 3D Objects." *Science*, vol. 347, no. 6228, 2015, pp. 1349-1352. 8. ISO/ASTM 52900:2021, *Additive Manufacturing: General Principles, Fundamentals and Vocabulary*. International Organization for Standardization, 2021. 9. GE Aerospace. "Transformation in 3D: How a Walnut-Sized Part Changed the Way GE Aviation Builds Jet Engines" and "Manufacturing Milestone: 30,000 (and later 100,000) Additive Fuel Nozzles." https://www.geaerospace.com/news 10. Bendsoe, M.P. and Kikuchi, N. "Generating Optimal Topologies in Structural Design Using a Homogenization Method." *Computer Methods in Applied Mechanics and Engineering*, vol. 71, 1988, pp. 197-224. 11. Autodesk / General Motors. "How GM and Autodesk Are Using Generative Design for Vehicles of the Future" (seat bracket: 40 percent lighter, 20 percent stronger, 8 parts consolidated to 1, 150+ candidate designs), 2018. https://adsknews.autodesk.com/en/news/gm-autodesk-using-generative-design-vehicles-future/ 12. See for example: Springer/ScienceDirect peer-reviewed studies on in-situ LPBF monitoring, including CNN porosity classification verified against X-ray CT (reported accuracy above 93 to 94 percent across multiple defect classes), *Journal of Intelligent Manufacturing* and *Optics and Laser Technology*, 2020-2024. 13. Nichol, Alex et al. "Point-E: A System for Generating 3D Point Clouds from Complex Prompts." OpenAI, arXiv:2212.08751, December 2022. https://openai.com/index/point-e/ 14. Jun, Heewoo and Nichol, Alex. "Shap-E: Generating Conditional 3D Implicit Functions." OpenAI, arXiv:2305.02463, May 2023. 15. Tencent Hunyuan3D Team. "Hunyuan3D 2.0: High-Resolution 3D Assets Generation with Large Scale Hunyuan3D Diffusion Models," January 2025. https://github.com/Tencent-Hunyuan/Hunyuan3D-2 16. Yahoo Finance / Meshy. "Meshy Hits $15M ARR with 30% Month-over-Month Growth," 2026. https://finance.yahoo.com/news/meshy-hits-15m-arr-30-032900666.html 17. PR Newswire. "Tripo AI Announces $50 Million in Funding and New Models for Production-Ready 3D Generation," March 2026. https://www.prnewswire.com/news-releases/tripo-ai-announces-50-million-in-funding-and-new-models-for-production-ready-3d-generation-302724894.html 18. Wehner, M. et al. "An Integrated Design and Fabrication Strategy for Entirely Soft, Autonomous Robots" (Octobot). *Nature*, vol. 536, 2016, pp. 451-455; Wyss Institute / Harvard SEAS press materials, August 2016. https://wyss.harvard.edu/news/the-first-autonomous-entirely-soft-robot/ 19. NPR / Spaceflight Now. "Relativity Space's 3D-printed Terran 1 launches, reaches space but falls short of orbit" (about 85 percent printed, launched March 22, 2023, Stargate printer). https://www.npr.org/2023/03/23/1162761566/3d-rocket-launch 20. FDA. *Technical Considerations for Additive Manufactured Medical Devices: Guidance for Industry and Food and Drug Administration Staff*, 2017 (updated 2023). 21. Ngo, Tuan D. et al. "Additive Manufacturing (3D Printing): A Review of Materials, Methods, Applications and Challenges." *Composites Part B: Engineering*, vol. 143, 2018, pp. 172-196. 22. Autodesk. "Generative Design in Fusion 360." Autodesk product documentation. 23. nTop. "Implicit Modeling for Engineering." nTop technical documentation. 24. Zoo (zoo.dev). "Text-to-CAD ML Documentation," 2024.