NOMAD: A Reusable Scientific Research Signifier

• NOMAD is a multi-domain research signifier used across materials science, planetary spectroscopy, optimization, machine learning, and biomedicine, each with distinct infrastructures and algorithms.
• In computational materials science, NOMAD underpins FAIR data infrastructures, open-access repositories, and AI toolkits that standardize and enable reproducible research.
• Other implementations span Mars atmospheric instruments, derivative-free optimization software, benchmark datasets for aerial vision and audio, and meta-taxonomies in biomedical nomenclature.

NOMAD is a recurrent acronym and project name in contemporary research literature rather than a single unified concept. Across materials science, planetary spectroscopy, optimization, machine learning, systems, software engineering, and biomedicine, the term denotes distinct infrastructures, instruments, algorithms, datasets, and experimental programs. In the cited literature, its most developed meanings include the Novel Materials Discovery data ecosystem in computational materials science, the Nadir and Occultation for MArs Discovery spectrometer on ExoMars Trace Gas Orbiter, derivative-free optimization software built around MADS, and several unrelated ML and data-analysis systems that independently reuse the name (Draxl et al., 2018, Liuzzi et al., 2018, Audet et al., 2021, Sridhar et al., 2023).

1. Scope of usage and recurrent expansions

The literature uses NOMAD both as an acronym with explicit expansions and as a retained software or project name whose meaning is domain-specific. A common misconception is to treat it as a single research lineage. The papers instead show repeated independent reuse across disciplines, with only some usages historically connected.

Usage	Domain	Role
Novel Materials Discovery	Computational materials science	FAIR data infrastructure, archive, analytics, and benchmarking (Draxl et al., 2018)
Nadir and Occultation for MArs Discovery	Planetary science	ExoMars TGO spectrometer suite for Martian atmosphere studies (Liuzzi et al., 2018)
Natural, Occluded, Multi-scale Aerial Dataset	Aerial computer vision	Benchmark for occluded human detection in SAR settings (Bernal et al., 2023)
Non-Matching Audio Distance	Speech and audio ML	Perceptual metric for non-matching-reference audio comparison (Ragano et al., 2023)
Nonlinear Manifold Decoders for Operator Learning	Neural operators	Nonlinear-decoder framework for function-space learning (Seidman et al., 2022)
Nomenclature Ontology for Medical And Disease names	Biomedical terminology	Meta-taxonomy for disease-name conventions (Denaxas et al., 11 Jun 2026)

This multiplicity is itself a salient feature of the term’s research usage. Some instances denote large-scale scientific infrastructures; others denote benchmark datasets, algorithmic frameworks, or single-domain software systems.

2. NOMAD in computational materials science

In computational materials science, NOMAD most prominently denotes the NOMAD Center of Excellence, a European infrastructure for making computational materials data genuinely big-data driven, open, and reusable. Its stated purpose is to collect, store, standardize, share, and enable reuse of original computational materials data, including raw input and output files. The infrastructure is organized around the NOMAD Repository, which stores raw code-produced data; the NOMAD Archive, which converts heterogeneous data into a common, code-independent format using NOMAD Meta Info; the NOMAD Encyclopedia, which exposes materials-oriented access to the archive; remote Visualization Tools including virtual reality; and an Analytics Toolkit for data mining and machine learning. The framework is explicitly aligned with the FAIR principles as “Findable,” “Accessible,” “Interoperable,” and “Re-purposable,” and the paper describes the Repository as the “world-wide largest raw-data collection of its kind” (Draxl et al., 2018).

The later NOMAD Artificial-Intelligence Toolkit extends this ecosystem into a browser-based environment for reproducible AI analysis. It operates on FAIR data from the NOMAD Archive as well as locally stored data and the local stand-alone NOMAD Oasis. The implementation is based on Docker, Jupyter notebooks, JupyterHub, and Kubernetes on MPCDF resources, with access through a unified RESTful API. The online environment provides up to 8 CPU cores and 10 GB RAM per user session, while the “Get to work” area provides 10 GB of cloud storage. The toolkit exposes feature libraries such as atomic_collections and matminer, supports workflows using HDBSCAN, t-SNE, and SISSO, and is designed so that users can share not only data but also versioned notebooks and methods under the Apache License Version 2 (Sbailò et al., 2022).

NOMAD also served as the organizing framework for the NOMAD 2018 Kaggle competition, a public benchmark on machine learning for materials-property prediction. The challenge used a dataset of 3,000 $(Al_xGa_yIn_z)_2O_3$ compounds with $x+y+z=1$, split into 2,400 training samples and 600 hidden test samples, and ran from December 18, 2017 to February 15, 2018 with 883 participants. Participants predicted electronic band gap energy and crystalline formation energy, with ranking based on the average RMSLE over both targets. The provided crystal structures were generated from pure binary geometries using Vegard’s law and were not fully relaxed, while the labels corresponded to fully relaxed geometries. The top three solutions were a crystal graph / n-gram representation + KRR winner, a descriptor-based model + LightGBM runner-up, and SOAP + neural network in third place. The subsequent nine-model comparison showed that, especially for formation energy, representation dominated more than the specific regressor, while low-correlation ensembles could outperform the best single model (Sutton et al., 2018).

3. NOMAD in Martian atmospheric spectroscopy

In planetary science, NOMAD denotes the Nadir and Occultation for MArs Discovery instrument suite aboard ESA’s ExoMars Trace Gas Orbiter. It was designed to observe Mars in solar occultation, nadir, and limb geometries, with particular emphasis on atmospheric composition and trace gases such as methane. The infrared channels combine an echelle grating spectrometer with an Acousto-Optical Tunable Filter (AOTF). Using first in-flight calibration data, the instrument paper reports residual calibration errors below 0.3 cm$^{-1}$, resolving power of about $R \sim 19{,}000$ for the SO channel and $R \sim 11{,}000$ for LNO, and a thermal sensitivity of about 0.75 pixel per °C. Under low aerosol conditions, the reported single-spectrum 1σ sensitivity to CH$_4$ in solar occultation is about 0.33 ppbv at 20 km altitude and better than 1 ppbv below 30 km; in dusty conditions the sensitivity drops to 0 below 10 km. In nadir geometry, seasonal methane maps are projected at around 5 ppbv sensitivity over most of the surface with 5° × 5° spatial bins, with binning yielding improvements by a factor of 10 to 30 (Liuzzi et al., 2018).

The retrieval framework Ares treats NOMAD SO as a physically modeled observation system for Bayesian inversion of Martian atmospheric composition. Ares extends TauREx 3 with a NOMAD SO instrument model, Martian geometry, molecular cross-sections, and a detector noise model. In this formulation, NOMAD comprises UVIS at 0.2–0.65 $\mu$m, SO at 2.3–4.3 $\mu$m, and LNO at 2.3–4.3 $\mu$m; SO uses diffraction orders 96–225, a detector with 320 spectral pixels and 256 spatial rows, and approximately 500 m vertical sampling over about 7.5 km using 24 detector rows. Ares models the coupled AOTF and blaze response, uses absorption data from HITRAN 2016, HAPI, and ExoCross, adopts Mars Climate Database priors for temperature, pressure, CO$_2$, and H$x+y+z=1$0O, and validates synthetic NOMAD spectra against the Planetary Spectrum Generator (Cann et al., 2020).

4. Optimization, numerical algorithms, and systems software

In optimization, NOMAD is a mature software package for blackbox optimization, especially nonlinear constrained and unconstrained problems for which derivatives are unavailable, unreliable, or impossible to compute. Its core method is MADS—Mesh Adaptive Direct Search—with a flexible search step and a convergence-critical poll step. NOMAD 4 is described as a complete redesign of version 3 around reusable algorithmic components, nested execution, and improved parallelism. New features highlighted in the paper include MegaSearchPoll, warm and hot restarts, a revised PSD-MADS, and parallel blackbox evaluation using OpenMP. The implementation is in C++14, uses CMake and Google Test, and runs on Linux and macOS. On a costly solar-plant benchmark, using 8 cores plus MegaSearchPoll yielded wall-clock times up to 3.3 times faster than a single-core configuration, while overall optimization performance was maintained between versions 3 and 4 (Audet et al., 2021).

A different NOMAD in scientific machine learning is Nonlinear Manifold Decoders for Operator Learning. This framework keeps the standard encoder–latent–decoder operator-learning template but replaces the usual linear decoder with a nonlinear decoder, motivated by the claim that many solution families to PDEs lie on low-dimensional nonlinear manifolds rather than low-dimensional linear subspaces. The paper shows order-of-magnitude gains over linear decoders on an antiderivative benchmark and lower-error representations on parametric advection. On the shallow-water benchmark, NOMAD achieved mean relative $x+y+z=1$1 errors of $x+y+z=1$2 for $x+y+z=1$3, $x+y+z=1$4 for $x+y+z=1$5, and $x+y+z=1$6 for $x+y+z=1$7, using $x+y+z=1$8 parameters, latent dimension 20, and about 5.5 min training cost, compared with larger baselines such as LOCA, DeepONet, and FNO (Seidman et al., 2022).

MD-NOMAD extends this operator-learning line to stochastic simulators by combining NOMAD with mixture density modeling. Instead of predicting a single output value, the decoder predicts parameters of a Gaussian mixture, allowing estimation of a conditional output distribution for SDEs and SPDEs. The framework is evaluated on the stochastic Van der Pol equation, Lorenz-96, an analytical bimodal density, and stochastic elliptic, heat, and Burgers equations. In the stochastic elliptic uncertainty-propagation example, MD-NOMAD reports relative $x+y+z=1$9 errors for mean and standard deviation of 0.0062 and 0.0251 for $^{-1}$0, and 0.0090 and 0.0455 for $^{-1}$1, improving substantially over deterministic NOMAD on standard-deviation estimation (Thakur et al., 2024).

Another computational use is the distributed matrix-completion algorithm NOMAD, expanded as Non-locking, stOchastic Multi-machine algorithm for Asynchronous and Decentralized matrix completion. It addresses low-rank matrix completion via asynchronous SGD in which user factors are partitioned permanently while item vectors are nomadic and transferred between workers. The system is decentralized, uses non-blocking communication, is lock-free, and is claimed to be serializable despite asynchrony. Empirically it outperformed synchronous methods such as DSGD, DSGD++, CCD++, and GraphLab variants on datasets including Netflix, Yahoo! Music, and Hugewiki, both on HPC clusters and on commodity AWS hardware (Yun et al., 2013).

In operating systems, Nomad is a Linux page-management framework for tiered memory that argues against exclusive page placement under memory pressure. Its core mechanisms are Transactional Page Migration (TPM) and page shadowing, which implement non-exclusive memory tiering by retaining a shadow copy of recently promoted pages in the slow tier. The paper reports up to 6x performance improvement over Linux TPP under memory pressure, while also analyzing scenarios in which hardware-assisted Memtis or even no migration may remain preferable (Xiang et al., 2024).

5. Machine learning policies, metrics, and benchmark datasets

In robotics, NoMaD—spelled with an internal capital M in the paper title—means Goal Masked Diffusion Policies for Navigation and Exploration. It is presented as a single policy that handles both goal-directed navigation and goal-agnostic exploration by combining a ViNT Transformer backbone, goal masking, and a diffusion decoder for action sequences. The model uses EfficientNet-B0 encoders, a 4-layer, 4-head Transformer decoder, and is trained on over 100 hours of real-world trajectories from GNM and SACSoN. On a real robot across 6 distinct indoor and outdoor environments, NoMaD achieved 98% exploration success with 0.2 collisions and 90% navigation success, matching the best navigation baseline while using 19M parameters rather than the 335M of the strongest prior Subgoal Diffusion baseline (Sridhar et al., 2023).

In aerial computer vision, NOMAD means Natural, Occluded, Multi-scale Aerial Dataset. It is a benchmark for human detection in emergency-response and search-and-rescue scenarios, explicitly centered on occlusion. The dataset contains 100 actors performing walking, laying, and hiding routines, with 42,825 manually annotated frames extracted from 5.4k videos at 5472 × 3078 and 30 fps. It uses 10 visibility levels and a five-distance acquisition setup at 10 m, 30 m, 50 m, 70 m, and 90 m. Benchmarking is performed with 10 folds and mAP@0.5:0.95 using YOLOv8l, FasterRCNN-R101-FPN, and RetinaNet-R101-FPN; the principal reported result is that detection performance drops sharply as visibility decreases and distance increases (Bernal et al., 2023).

In speech and audio, NOMAD stands for Non-Matching Audio Distance, a differentiable perceptual similarity metric for comparing degraded speech to non-matching clean references. The method builds a 256-dimensional embedding on top of wav2vec 2.0 BASE, trained by triplet loss with NSIM guidance. It is evaluated on degradation ranking, speech-quality prediction, and speech-enhancement loss design. The paper reports that NOMAD outperforms prior non-matching-reference approaches, remains competitive with full-reference metrics, and improves DEMUCS when used as a perceptual loss in both fine-tuning and multitask settings (Ragano et al., 2023).

A separate ML and visualization use is NOMAD Projection, expanded as Negative Or Mean Affinity Discrimination. This is a nonlinear dimensionality-reduction method designed to scale across multiple GPUs at train time, and is positioned as an approximate upper bound on the InfoNC-t-SNE loss. Its principal algorithmic idea is to keep positive forces local through a shardable ANN graph and to approximate many negative forces using cluster means. On the PubMed benchmark, the paper reports NP@10 = 6.1 ± 0.3% in 1.47 hours on 8× H100 GPUs, compared with OpenTSNE at NP@10 = 6.2% in 8.0 hours, while RapidsUMAP and t-SNE-CUDA ran out of memory. The method also produced the first complete map of Multilingual Wikipedia, comprising 60 million points, in about 5.8 hours on an 8× H100 node (Duderstadt et al., 21 May 2025).

6. Additional domain-specific systems and experimental programs

In high-energy physics, NOMAD refers to the neutrino experiment whose full data set enabled a precision measurement of charm dimuon production in neutrino–iron interactions. The paper reports 15,344 background-subtracted charm dimuon events from a much larger inclusive charged-current sample of about $^{-1}$2 events after cuts, yielding an average dimuon-to-CC ratio of $^{-1}$3. An NLO QCD fit gives a running charm mass of $^{-1}$4 and a strange-sea suppression factor of $^{-1}$5 at $^{-1}$6 (Collaboration, 2013).

In software engineering, NOMAD is also the name of a modular multi-agent LLM framework for generating UML class diagrams from natural-language requirements. Its pipeline comprises four generation agents—Concept Extractor, Relationship Comprehender, Model Integrator, and Code Articulator—plus a Validator. Evaluation uses the Northwind 2.0 case study and eight human-authored UML exercises. On the supplementary exercises, the reported average F1 improves from 0.6561 for a single-agent baseline to 0.6958 for NOMAD; on Northwind, average F1 rises from 0.7360 to 0.8360, driven especially by stronger relationship extraction, although attribute extraction remains difficult. With GPT-4o, the verifier further raises average F1 from 0.8360 to 0.8850 (Giannouris et al., 27 Nov 2025).

For real-time stream processing, NOMAD—Navigating Optimal Model Application to Datastreams—is a framework that constructs per-event model chains over pretrained classifiers with different cost-quality tradeoffs. It uses a utility criterion inspired by predicate ordering in database systems, defines exit classes relative to a designated role model, and enforces chain safety so that quality remains $^{-1}$7-comparable to that role model. Across eight datasets, the paper reports normalized F1 between 0.91 and 1.03, speedups from 2.1× to 6.4×, and, on UNSW-NB15, throughput of 1494 events/sec compared with 457 events/sec for always using the role model (Colaco et al., 31 Oct 2025).

In autonomous research systems, Nomad denotes an exploration-first architecture for discovery over corpora, databases, and other sources. Its core data structure is an explicit Exploration Map that organizes topics, concepts, hypotheses, insights, and documents; it then uses topic selection, hypothesis generation, an explorer–verifier loop, and a reporting pipeline producing cited reports and meta-reports. On selected UN and WHO corpora, the system is reported to achieve higher trustworthiness and quality than GPT Researcher and o3-deep-research, while also exhibiting stronger diversity across runs; for example, bottom-up inter-report diversity is 0.4273, and top-down diversity is 0.4697 (Jia et al., 31 Mar 2026).

In biomedical terminology, NOMAD means Nomenclature Ontology for Medical And Disease names. It is a meta-taxonomy of disease-name construction with 9 top-level categories and 20 subcategories, applied to 22,548 entries from the ICD-10-CM 2026 Alphabetical Index in a three-stage classification pipeline. The study classifies 99.1% of entries with a mean of 2.12 labels per entry, finds Anatomical labels in 63.8% of entries, Descriptive in 48.4%, and Pathophysiological in 40.2%, and reports a macro-averaged Cohen’s $^{-1}$8 of 0.832 on a 10% manual review sample. The resulting picture is that disease names are typically multi-label and compositional rather than governed by a single naming principle (Denaxas et al., 11 Jun 2026).

Taken together, these usages show that NOMAD functions in the literature as a highly reusable scientific signifier rather than a stable cross-domain brand. Its meanings range from data stewardship and atmospheric instrumentation to derivative-free optimization, neural operator design, ML benchmarking, autonomous discovery, and experimental physics. The commonality lies not in shared method or lineage, but in the repeated selection of the same short, memorable name for technically unrelated research artifacts.