NIST AM Bench Additive Manufacturing Benchmark
New1. Overview
Simulation models in additive manufacturing face a structural difficulty. The ground truth for melt pool temperature fields, residual stress distributions, and microstructure morphology is extremely hard to obtain. In laser powder bed fusion the process happens inside a melt pool a fraction of a millimetre across, cooling rates reach thousands of degrees per second, and the finished component retains residual strain. The only way to validate a process simulation model is to find high-accuracy measurements of the same specimen, and such data has traditionally been a private by-product of individual papers that nobody else can re-examine.
AM Bench exists to make that validation chain public. Led by the U.S. National Institute of Standards and Technology, it produces measurements with annotated uncertainty through standardized artifacts, controlled additive manufacturing equipment, and rigorous metrology, then issues blind challenge problems that the modelling community answers before the measurements are revealed. The mechanism turns a claim that a model looks good into a statement of how far that model sits from a published gold standard on a neutral benchmark.
The relevance to physical AI is that it offers a different evaluation paradigm. Physical AI work relies heavily on validation splits and metric comparison, which typically capture average performance inside the data distribution. The blind challenge is far stricter: predictions are submitted before measurements are released, and the whole set is evaluated afterwards, so the distribution of errors rather than its mean becomes the primary quantity. The method discussion in the surrogate modeling and reduced-order model collection reads more clearly against that background, since a surrogate accuracy claim needs exactly this kind of external yardstick to stand on.
On the engineering side, AM Bench has moved beyond simple test specimens toward validating higher technology readiness level applications, which is stated explicitly in the 2026 planning.
2. How It Works
The organizational form is a NIST-led body that provides a continuing series of benchmark measurements, challenge problems, and conferences, with the primary goal of enabling modelers to test their simulations against rigorous and highly controlled additive manufacturing benchmark measurement data. All AM Bench data is permanently archived for public use through comprehensive custom data management systems, and the data site is built on the Materials Data Curation System from NIST.
The cycle structure runs as follows. The first round of benchmark measurements, challenge problems, and conference, AM Bench 2018, was completed in 2018. The second round, AM Bench 2022, was completed after a one-year delay caused by the COVID-19 pandemic. All work on AM Bench 2025 has been completed except for finishing the final data publications and measurement papers. Planning for AM Bench 2028 is underway, and additional asynchronous benchmarks are also being planned.
Conferences follow a three-year cycle. The first AM Bench conference, AM Bench 2018, was hosted by NIST at the Gaithersburg, Maryland campus during the week of 18 June 2018. The second, AM Bench 2022, was held in Bethesda, Maryland during the week of 15 August 2022.
Artifact selection follows a stated principle. AM Bench was established to develop benchmark measurements for all AM processes and materials, but the requirement for highly controlled and quantitative measurements severely limits the number of benchmark measurements that can be conducted in any given test cycle. The key challenge is therefore to select measurements, processes, and materials that are both technically feasible and have the highest impact on the AM community, with the AM Bench Scientific Committee advising on the final choices used for the 2018, 2022, and 2025 benchmarks.
Measurement techniques divide into in-situ and post-build characterization. In-situ means include melt pool imaging, thermography, and in-situ synchrotron diffraction for capturing phase transformation at the moment of solidification. Post-build means include X-ray CT residual strain reconstruction, neutron diffraction, synchrotron X-ray diffraction, contour method part deflection measurement, and EBSD microstructure mapping. Residual strain alone is addressed with three independent techniques: neutron diffraction, synchrotron X-ray diffraction, and the contour method.
The blind challenge mechanism is central. The benchmarks include blind challenge problems to the AM modeling community where modelers are asked to predict the AM Bench measurement results. Predictions are submitted first, and the measurement data is released only after the submission deadline, making the error a comparable public record.
The 2018 scope covered metals and polymers. For metals the focus was laser powder bed fusion of a nickel-based superalloy, 625, and a martensitic precipitation-hardenable stainless steel, 15-5, with both three-dimensional builds and single laser tracks on bare metal plates conducted. For polymers the focus was thermoplastic technologies, with materials extrusion and LPBF, also referred to as selective laser sintering, identified as the two most widely used approaches. The MatEx benchmark focused on polycarbonate and the LPBF benchmark on polyamide 12.
The 2022 scope continued both tracks. For metals, three sets of benchmarks focused on LPBF of nickel-based superalloy 718 and two sets provided follow-on mechanical performance and microstructure measurements for the 2018 PBF-LB studies on alloy 625. Taken together these cover the full processing-structure-properties range including feedstock characterization, in-situ measurements during the build, heat treatments, two- and three-dimensional microstructure characterization, residual strains and part deflection, and mechanical behavior measurements. For polymers, one set focused on material extrusion of polycarbonate and another on vat photopolymerization measurements. An additional asynchronous set completed in early 2022 focused on melt pool and laser coupling dynamics during high-power laser interactions with titanium-alloy and aluminum-alloy bare metal surfaces.
The 2025 scope comprises eight sets of measurements and challenge problems for metals and one set for polymers. The metals benchmarks included several new focus areas: composition sensitivity with one set, high-cycle fatigue with one set, laser hot-wire DED with two sets, and time-resolved measurements of phase transformations immediately following solidification with one set. Another metals set was a mechanical properties follow-on to an AM Bench 2022 set of PBF-LB benchmarks, and two additional sets continued the tradition of measuring the underlying physical processes for PBF-LB. The polymers set focused on cure depth for vat photopolymerization.
3. Scope and Assumptions
Access to the data has a threshold. The data is hosted on the CDCS system, where browsing, searching, and downloading require an account, and some content additionally requires the custom query composer. API documentation is provided separately.
Measurement coverage is bounded by technical feasibility. The official statement notes that the requirement for highly controlled and quantitative measurements severely limits the number of benchmark measurements that can be conducted in any given test cycle, which means some process and material combinations simply have no benchmark in a given round.
The three-year cycle carries a waiting cost. All work on AM Bench 2025 is complete except the final data publications and measurement papers, and AM Bench 2028 is still in planning. Projects treating AM Bench data as a training set must weigh data freshness and version updates.
Process and material coverage has edges. The combinations covered concentrate on LPBF of nickel-based superalloys 625 and 718, martensitic precipitation-hardenable stainless steel 15-5, polymers in MatEx and vat photopolymerization, and laser hot-wire DED. Aluminum and titanium alloys appear only in the asynchronous benchmark as bare-surface laser interaction, without complete component-level mechanical data.
Not every physical quantity has a benchmark. Coverage is a selected set of measurement items along the processing-structure-performance chain, and areas such as multiphase flow, powder spreading dynamics, support strategy, and defect formation mechanisms still lack public benchmarks of comparable strength.
Blind predictions mean some data already has answers. Challenge measurements are public after the deadline, so training a model on them and evaluating on the same set is data leakage and is not accepted academically.
The 2028 shift to higher TRL changes the nature of the data. Official statements note that model maturity now allows a shift from simple specimens to realistic metal components designed with industry, with the first such benchmarks planned for AM Bench 2028. Later data will be closer to industrial parts but cover fewer controlled variables, and that trade-off must be flagged when citing across rounds.
4. Getting Started
Register and learn the data portal. Create an account on ambench.nist.gov and browse through the Materials Data Curation System. The home page offers four paths: data curation, keyword search, example queries, and the custom query composer, with separate API documentation.
Read the measurement report before downloading data. Each benchmark measurement paper contains the method and uncertainty budget, and NIST explicitly asks users to cite the relevant AM Bench publication. Only by comparing against the uncertainty budget can you tell whether an error magnitude falls inside measurement noise.
Pick a matching challenge. Challenges are organized by physical quantity with a numbering scheme. Examples include AMB2022-01-MS for microstructure, CHAL-AMB2022-04-MeTT for the subcontinuum mesoscale tensile challenge, AMB2018-03 for MatEx polymer physical properties, and AMB2018-04 for LPBF polymer physical properties.
Run your own blind test. Take the geometry, material, and process parameters of a measurement artifact, submit predictions, and then compare against the published measurements. Completing this loop once yields a precision coordinate for your own solver on that process route.
Evaluate uncertainty propagation. Residual strain is measured with three independent techniques that cross-validate one another, so predictions should be compared against an uncertainty band rather than a point value.
Choose a modelling route. For fast surrogate models, the AFLOW dataset entry is a reference. For embedding physics into neural networks, DeepXDE and PhysicsNeMo provide implementation paths for physics-constrained networks. For multiscale structure prediction, graph neural network frameworks such as PyTorch Geometric are a closer fit.
Watch the direction of later rounds. The higher TRL focus for AM Bench 2028 and the planned expansion to additional AM methods and materials are described on dedicated pages on the portal and are worth factoring into project planning early.
5. Comparison with Alternatives
Against an in-house benchmark. Company-internal artifacts with internal metrology are the most common alternative, offering faster iteration and freedom from public release schedules, but measurement uncertainty is rarely independently annotated, third parties cannot re-examine the data, and the honesty constraint that blind testing provides is absent. The value of AM Bench concentrates precisely on that publicly verifiable property.
Against the AFRL additive manufacturing modeling challenges. The Air Force Research Laboratory series is a comparable mechanism under a different organizer and material set. Published AFRL challenge papers in Integrating Materials and Manufacturing Innovation cover Challenge 1 on residual strain distribution characterization, Challenge 2 on microscale process-to-structure data description, and an overview paper, with materials concentrated on IN625 and AFRL systems. The two can be cross-referenced, and matching by material system saves effort when choosing topics.
Against single-paper dataset releases. The usual practice is a paper shipping its own test data, which is tightly coupled to a specific model, lacks third-party cross-model comparison, and often omits uncertainty information.
Against built-in validation in commercial additive software. Vendor example cases serve functional regression, which is a different activity from physical accuracy validation. The latter needs an external gold standard.
Against online process monitoring data. In-situ melt pool monitoring yields large continuous streams but provides process signals rather than offline high-accuracy measurements, so it lacks a gold standard anchor. The two are complementary: online data supports modelling and control, and models calibrated online are then graded against a benchmark such as AM Bench.
Against AI-driven process optimization work. Research applying artificial intelligence and statistical mapping to additive manufacturing toolpath optimization shows that benchmark data is already feeding AI methods, and the corresponding evaluation framing belongs with the surrogate modeling and reduced-order model discussion.
6. Benchmarks and Evidence
Challenge participation in 2025 is recorded as 85 submissions to AM Bench challenge problems from the modeling community.
Data usage figures are striking. The AM Bench 2022 microstructure challenge AMB2022-01-MS received 3 submissions, yet as of 21 July 2026 NIST tracked 6044 unique-user downloads of the corresponding data publication. Across all published AM Bench data publications, approximately 1670 unique user downloads per month, or 55 per day, were tracked between 2 May 2025 and 1 May 2026. This comparison shows that the reach of the datasets far exceeds the participation in any single challenge.
The 2022 benchmark set consists of three metals sets focused on LPBF of IN718, two follow-on IN625 mechanical and microstructure sets, two polymers sets covering polycarbonate MatEx and vat photopolymerization, and one asynchronous set completed in early 2022 on melt pool and laser coupling dynamics for high-power laser interaction with titanium and aluminum alloy bare plates.
The 2025 benchmark set consists of eight metals sets, namely composition sensitivity, high-cycle fatigue, two laser hot-wire DED sets, time-resolved phase transformation measurements immediately after solidification, a mechanical properties follow-on to the 2022 LPBF benchmarks, and two sets measuring the underlying physical processes of LPBF, plus one polymers set on cure depth for vat photopolymerization.
Published test conditions for 2025 are recorded in official image captions, covering the time above melting for arrays of laser tracks on nickel alloy 718 bare plate at a laser turn-around time of 0.75 ms, an in-situ diffraction pattern acquired during solidification of an Fe-Cr-Ni alloy melt pool, and a laser hot-wire DED three-dimensional build for AM Bench 2025.
Representative published measurement papers include location-specific microstructure characterization within IN625 AM-Bench artifacts in volume 9 pages 54-69 of March 2020, elastic residual strain and stress measurements and corresponding part deflections for IN625 AM-Bench artifacts using neutron diffraction, synchrotron X-ray diffraction, and the contour method in volume 8 pages 318-334 of July 2019, a study on melt pool shape prediction accuracy using the non-linear temperature equation including phase changes in volume 8 pages 167-177 of April 2019, AMB2018-04 polymer LPBF physical property measurements in volume 8 pages 335-361, AMB2018-03 polymer MatEx physical property measurements in volume 9 pages 358-375, part deflection measurements of AM-Bench IN718 three-dimensional build artifacts in volume 12 pages 386-396 of October 2023, and the CHAL-AMB2022-04-MeTT subcontinuum mesoscale tensile challenge with prediction summary in volume 12 pages 196-209 of September 2023.
The AFLOW machine learning benchmark datasets were published in volume 9 pages 153-156 of May 2020, showing that this benchmark data has already been packaged in a directly trainable form.
Conference and publication details: the 2018 proceedings appeared in the TMS journal Integrating Materials and Manufacturing Innovation within the thematic section Additive Manufacturing Benchmarks 2018, containing both AM Bench measurement results and invited papers from the modeling community, and the 2022 proceedings were also published through IMMI. Due to a change in policy TMS no longer supports small independent conferences such as AM Bench, so AM Bench 2025 was held in conjunction with ASME's International Mechanical Engineering Congress and Exposition as a separate track at the Renasant Convention Center in Memphis, Tennessee on 16-20 November 2025, with the 2025 proceedings again to be published in IMMI. The NIST organizing committee co-chairs are Lyle E. Levine and Jordan Weaver.
7. Role in Physics-AI Workflows
One use is providing an external yardstick for additive process surrogate models. The most common failure of a surrogate model is good in-distribution behaviour and degradation on new conditions. The blind challenge provides a cross-condition public reference point, so the question of what accuracy level a surrogate actually reaches has a citable answer.
A second use is training uncertainty-aware models. Because benchmark data carries uncertainty annotation, joint modelling of prediction error and measurement uncertainty becomes possible, avoiding the mistake of reading measurement noise as model error.
A third use is supplying validation sets for physics-constrained networks. Models that write melt pool heat transfer, latent heat of phase transformation, and residual stress relaxation into the loss function need their extrapolation ability tested against real measurements. The validation difficulties for this class of method are discussed in the PINN landscape guide, and AM Bench provides ready-made comparison data.
A fourth use is providing an anchor point for multiscale models. Bridging from melt pool scale to component scale needs a commonly accepted anchor, and benchmark data supplies third-party reproducible numbers at that interface.
A fifth use is the higher TRL validation path for industry. The 2028 pivot toward real metal components designed with industry, such as the laser powder bed fusion jet engine stator segment shown in the official imagery, yields benchmarks directly usable by industry to test predictive capability on production-grade parts.
A sixth use is feeding benchmark data into the materials digital thread. Complete-chain measurements from feedstock characterization to mechanical performance provide high-quality samples for building materials informatics data assets, which is precisely the kind of source the physics foundation model collection discusses.
8. Limitations and Failure Modes
It cannot support process development decisions. Benchmarks measure physical outcomes of an existing process and contain no causal relationship between process parameters and performance. Determining whether a parameter window is better requires a dedicated experimental design.
It does not span the whole computational space. Only certain process and material combinations are covered, and calculations outside those combinations can use the data as a reference value but not as an acceptance benchmark.
Cycle lag affects timeliness. On a three-year cadence, new processes often have no matching benchmark. The asynchronous benchmark completed in early 2022 covers part of the high-power laser surface interaction scenario.
Blind results involve selective submission. The official count of 85 submissions in 2025 is not a measure of method coverage, so it cannot be used to judge which method families were systematically tested.
Download volume is not depth of use. High download counts signal attention and equally signal the risk of repeated metric chasing on the same data, so citations should state the intended use.
The higher TRL pivot changes data character. Once the 2028 data turns to real industrial components with fewer controlled variables, it is not homogeneous with the 2018 to 2025 specimen-level data, and direct cross-round comparison needs care.
Measurement uncertainty is a hard constraint. The feasibility of measuring some quantities is itself contested, illustrated by a dedicated paper on benchmark sample design for validating residual stress measurements by diffraction. Predictions must align with the uncertainty band, and point-by-point comparison yields misleading conclusions.
Defect and failure mechanisms are not covered. Porosity, cracking, and interlayer bonding lack public benchmarks of comparable strength.
9. Version and Source Notes
On round status. According to the current official portal: AM Bench 2018 was completed in 2018, AM Bench 2022 in 2022 after a one-year COVID-19 delay, all work on AM Bench 2025 is complete except the final data publications and measurement papers, and planning for AM Bench 2028 is underway with additional asynchronous benchmarks planned. The portal also states that a key change expected for AM Bench 2028 is a new focus on higher technology readiness level applications while continuing to support the lower TRL base.
On the data site. The platform is ambench.nist.gov backed by the Materials Data Curation System at cdcs.nist.gov, with the data management system identified as provided at the usnistgov/mdcs repository. Data is permanently archived for public use. The site exposes the CDCS home page, data curation, keyword search, example queries, the custom query composer, API documentation, contact, and help, and links to the AM Bench topical collection in the IMMI journal. The data portal currently reports Materials Data Curation System version 3.14.0.
On licensing. The official wording is that all AM Bench data is permanently archived for public use, and the data itself is a U.S. NIST public-domain product. The license and licenseEn fields of this entry have been rewritten accordingly and no longer read generically as referencing a LICENSE file.
On data statistics timestamps. The download figures use different windows and must be labeled separately: a cumulative count as of 21 July 2026, and an annual interval covering 2 May 2025 to 1 May 2026.
On visual asset attribution. Official imagery on the portal is credited to NIST, with photography by Matthew Dantin of the Naval Surface Warfare Center Carderock Division, and Jordan Weaver of NIST. The jet engine stator segment image corresponds to the AM Bench 2028 plan. The page notes that one visual asset was created with generative super intelligence.
On proceedings publication. The 2018 and 2022 conference proceedings were published in IMMI by TMS, and the 2025 proceedings will again be published in IMMI, though the 2025 conference was held jointly with the ASME IMECE because of the TMS policy change.
On challenge numbering. Known identifiers include AMB2018-03 for MatEx polymer physical properties, AMB2018-04 for LPBF polymer physical properties, AMB2022-01-MS for the microstructure challenge, and CHAL-AMB2022-04-MeTT for the subcontinuum mesoscale tensile challenge. Full identifiers should be used when citing specific measurement items.
On related methodology papers. A common data dictionary and common data model for additive manufacturing appeared in volume 13 pages 105-119 of February 2024, providing a standardized framing for cross-source data integration, and a paper on benchmark sample design for the validation of residual stress measurements by diffraction appeared in volume 13 pages 955-968 of November 2024.
10. References
- U.S. National Institute of Standards and Technology, official AM Bench portal covering round status, the 2018, 2022, and 2025 benchmark scopes, challenge submission statistics, dataset download statistics, and conference and publication arrangements, https://www.nist.gov/ambench
- AM Bench Materials Data Curation System data site with data curation, keyword search, custom query composer, and API documentation, https://ambench.nist.gov/
- Official NIST image captions for the AM Bench 2018 build plate, the 2025 laser scan track time-above-melting distribution, and the in-situ diffraction pattern of an Fe-Cr-Ni melt pool, referenced from the AM Bench official portal, https://www.nist.gov/ambench
- Phan L., Strantza E., Hill M., Gnaupel-Herold M., Heigel J., et al., "Elastic Residual Strain and Stress Measurements and Corresponding Part Deflections of 3D Additive Manufacturing Builds of IN625 AM-Bench Artifacts Using Neutron Diffraction, Synchrotron X-Ray Diffraction, and Contour Method", Integrating Materials and Manufacturing Innovation, Vol. 8, July 2019, pp. 318-334, DOI 10.1007/s40192-019-00149-0
- Stoudt M., Williams J., Levine L., Creuziger E., Young T., Heigel J., et al., "Location-Specific Microstructure Characterization Within IN625 Additive Manufacturing Benchmark Test Artifacts", Integrating Materials and Manufacturing Innovation, Vol. 9, March 2020, pp. 54-69, DOI 10.1007/s40192-020-00172-6
- Kafka A., Benzing A., Moser E., Liew K., Weaver J., Hrabe T., "Additive Manufacturing Benchmark 2022 Subcontinuum Mesoscale Tensile Challenge (CHAL-AMB2022-04-MeTT) and Summary of Predictions", Integrating Materials and Manufacturing Innovation, Vol. 12, September 2023, pp. 196-209, DOI 10.1007/s40192-023-00307-5
- Praniewicz M., Fox J., Tarr J., "Part Deflection Measurements of AM-Bench IN718 3D Build Artifacts", Integrating Materials and Manufacturing Innovation, Vol. 12, October 2023, pp. 386-396, DOI 10.1007/s40192-023-00310-w
- Clement J. A., Kauwe J., Sparks M., "Benchmark AFLOW Data Sets for Machine Learning", Integrating Materials and Manufacturing Innovation, Vol. 9, May 2020, pp. 153-156, DOI 10.1007/s40192-020-00174-4
- Kollmannsberger S., Carraturo M., Reali A., Auricchio C., "Accurate Prediction of Melt Pool Shapes in Laser Powder Bed Fusion by the Non-Linear Temperature Equation Including Phase Changes", Integrating Materials and Manufacturing Innovation, Vol. 8, April 2019, pp. 167-177, DOI 10.1007/s40192-019-00132-9
- "The AFRL Additive Manufacturing Modeling Challenge Series: Overview", Integrating Materials and Manufacturing Innovation, Vol. 10, June 2021, pp. 125-128, DOI 10.1007/s40192-021-00215-6
- "A Common Data Dictionary and Common Data Model for Additive Manufacturing", Integrating Materials and Manufacturing Innovation, Vol. 13, February 2024, pp. 105-119, DOI 10.1007/s40192-024-00341-x
- U.S. NIST Materials Data Curation System open-source repository backing the data portal, https://github.com/usnistgov/mdcs