DALIA: Cross-Domain Research Frameworks

• DALIA is a cross-domain term that denotes distinct research assets, ranging from a chemistry OER discovery platform to a wearable sensing benchmark, an HIV vaccine trial, an agentic layer, and a scalable Bayesian framework.
• It enhances research workflows through rich metadata models, standardized benchmarks, sequential nonparametric clustering, and declarative planning methodologies across diverse applications.
• Its scalable INLA framework accelerates multivariate spatio-temporal Bayesian inference, achieving remarkable computational performance on high-performance computing architectures.

In the cited research literature, DALIA is not a single object but a cross-domain designation applied to several distinct infrastructures, datasets, and study contexts. The name appears in chemistry education as an open educational resources discovery platform for research data management training, in wearable sensing as PPG-DaLiA, in HIV immunology as the DALIA-1 therapeutic-vaccine trial, in LLM systems as the Declarative Agentic Layer for Intelligent Agents, and in scientific computing as a scalable framework for Bayesian inference in multivariate spatio-temporal Gaussian processes (Hoffmann et al., 23 Sep 2025, Kasnesis et al., 2022, Hejblum et al., 2017, Rodriguez-Sanchez et al., 24 Jan 2026, Gaedke-Merzhäuser et al., 9 Jul 2025).

1. Referential scope and naming

The term appears with several capitalization patterns. DALIA is used for the chemistry platform, the declarative agentic layer, the HIV trial designation, and the high-performance Gaussian-process framework; DaLiA appears in PPG-DaLiA, the wearable physiological dataset and benchmark family.

Referent	Domain	Description
DALIA OER Platform	Chemistry education	OER discovery platform supporting FAIR-oriented RDM teaching alongside Chemotion (Hoffmann et al., 23 Sep 2025)
PPG-DaLiA	Wearable sensing	Wrist/chest multimodal dataset collected in close-to-real-life conditions (Kasnesis et al., 2022)
DALIA-1	HIV immunology	Phase I therapeutic vaccine trial against HIV (Hejblum et al., 2017)
Declarative Agentic Layer for Intelligent Agents	Agent systems	Declarative layer for MCP-based server ecosystems (Rodriguez-Sanchez et al., 24 Jan 2026)
DALIA	Bayesian HPC	INLA-based framework for multivariate spatio-temporal GPs (Gaedke-Merzhäuser et al., 9 Jul 2025)

This multiplicity is not merely terminological. The cited works attach the name to materially different objects: a metadata and curation layer for educational resources, a benchmark corpus for physiological ML, a longitudinal clinical study, an execution architecture for tool-using agents, and a GPU-centric inference stack. A plausible implication is that any technical use of “DALIA” requires immediate domain qualification.

2. DALIA as a chemistry OER discovery platform

In chemistry education, DALIA is presented as an open educational resources discovery platform that complements, rather than replaces, the practical research-data workflow centered on the Chemotion electronic laboratory notebook. In the RWTH Aachen University fifth-semester advanced inorganic laboratory course, students use Chemotion to plan, document, and evaluate experiments digitally, while DALIA serves as the “discovery layer” through which students and teachers find chemistry-specific RDM learning materials, especially tutorials for Chemotion and FAIR-oriented data practices (Hoffmann et al., 23 Sep 2025).

The platform is explicitly not a laboratory notebook and not the system into which experimental chemistry data are recorded. Its function is to index and expose educational resources using a richer metadata model than generic repositories commonly provide. The paper identifies missing learner-centered descriptors in many OER repositories—such as resource type, media type, discipline, target group, and learner proficiency level—as a major barrier to educational reuse. DALIA addresses that barrier through the DALIA Interchange Format (DIF), which includes descriptive fields such as title, description, language, keywords, license, external URL, date published, and author list, and extends them with controlled classifications for resource type, media type, discipline, target group, and learner proficiency level (Hoffmann et al., 23 Sep 2025).

The technical emphasis is interoperability. DALIA’s metadata model is implemented with RDF and reuses the Schema.org educational resource data model, extending it where required for chemistry learning. The platform indexes external OERs rather than hosting all materials locally, links them to communities and collections, and exposes a faceted search interface intended to let, for example, “a bachelor’s level chemist” identify a chemistry tutorial appropriate for beginners. The Chemotion instructional videos are the paper’s main case study: DALIA enriches a Zenodo-hosted resource with community, discipline, target-group, and proficiency metadata, thereby improving pedagogical findability (Hoffmann et al., 23 Sep 2025).

The educational corpus described in the paper comprises 405 OERs. Of these, 172 of 405 resources (42.5%) are targeted toward undergraduate bachelor students, and 180 of 405 (44.4%) toward beginners. The paper presents these figures as evidence that the platform is oriented toward early-stage learners rather than only expert audiences. At the same time, the authors identify metadata quality as the main operational bottleneck: discipline, keyword, and proficiency annotations remain inconsistent, motivating semi-automated enrichment, community curation, and future harvesting from repositories such as Zenodo and DARIAH-Campus (Hoffmann et al., 23 Sep 2025).

3. PPG-DaLiA as a wearable physiological benchmark

PPG-DaLiA is a public multimodal dataset for wrist-based physiological inference under realistic movement. One widely used description specifies 15 subjects aged 21–55, a wrist-worn device on the non-dominant wrist, 1-channel PPG at 64 Hz, 3-axis accelerometer at 32 Hz, and ECG-derived ground truth from a chest sensor (Kasnesis et al., 2022). Other papers describe the corpus as containing about 36 hours of data (Basso et al., 14 Aug 2025) or 37.5 hours of recording across 8 daily activities (Burrello et al., 2022), reflecting slightly different summaries of the same resource.

The benchmark’s difficulty derives from motion artifacts. Activities such as walking, transitions, and stair ascent or descent introduce nonstationary contamination caused by sensor displacement and related effects, so the dataset has become a standard testbed for wrist-PPG robustness claims (Kasnesis et al., 2022). Depending on task definition, papers use different task-specific views of the corpus. For heart-rate estimation, a common setup segments the data into 8-second windows; one influential formulation downsampled all channels to 32 Hz, obtained an input tensor of size $4 \times 256$, and used per-channel z-score normalization under subject-independent evaluation (Kasnesis et al., 2022). Other work preserved the original 64 Hz PPG and resampled later according to model design (Geenjaar et al., 5 Jun 2026).

The dataset is not limited to wrist PPG and accelerometry. The QXAI study describes a broader feature space with 11 attributes spanning physiological and motion measurements, including 3D acceleration, ECG, respiration, BVP, EDA, and body temperature (Shaik et al., 2023). A forecasting adaptation later transformed the resource into a 15-channel multivariate sequence by combining chest and wrist modalities plus interpolated heart-rate labels, resampling to 1 Hz, concatenating subjects chronologically, and yielding 52,696 timesteps under a 70/10/20 split that was explicitly not subject-disjoint (Mysore, 27 Apr 2026).

This task dependence is central to PPG-DaLiA’s role in the literature. The same underlying recordings support subject-independent HR regression, RR estimation, motion-artifact segmentation, IBI and HRV recovery, activity classification, federated personalization experiments, reinforcement-learning monitoring environments, and multivariate forecasting. This suggests that PPG-DaLiA functions less as a single benchmark protocol than as a shared physiological substrate on top of which multiple, sometimes non-comparable, evaluation regimes are defined.

4. Methodological ecosystem built around PPG-DaLiA

The best-known PPG-DaLiA use case is heart-rate estimation from wrist PPG plus motion. In a subject-independent LOSO protocol, Q-PPG reported 4.36 BPM MAE with post-processing (Burrello et al., 2022). PULSE then reduced this to 4.03 BPM MAE, a 7.56% relative reduction over Q-PPG*, by combining temporal convolutions with 4-head cross-attention in which PPG acts as query and accelerometer as key/value (Kasnesis et al., 2022). EnhancePPG later improved the best reported error from 4.03 BPM to 3.54 BPM by applying self-supervised reconstruction pretraining and augmentation to a modified PULSE backbone, while keeping inference unchanged relative to that modified backbone (&&&24&&&). More recently, PCHS reported 3.20 bpm MAE for HR and 2.15 bpm MAE for 32 s aggregated RR, using a harmonic physiological decoder, a motion-conditioned residual branch, and uncertainty-aware weighting (Fraihi et al., 29 Jun 2026).

Comparisons across papers are not protocol-invariant. The QXAI paper, for example, used an 80:20 split over a prepared dataset rather than subject-wise evaluation and reported MAE 3.33 and MSE 24.51 for an ANN, with SHAP- and attention-based explanation layers used to attribute contributions from activity, chest and wrist motion, temperature, and cardiac features (Shaik et al., 2023). That result is therefore informative about explainability under the authors’ setup, but it is not directly commensurate with the LOSO literature.

PPG-DaLiA has also supported several adjacent methodological directions. PCE-LSTM defined “DaLiA” as the PPG-DaLiA subset without PPG for IMU-only multi-step HR prediction, using a learned Physical Conditioning Embedding to initialize the recurrent state; on this task it achieved ensemble MAE 13.5, compared with 15.9 for FFNN and 16.5 for DeepConvLSTM, and on the full PPG-DaLiA task it reported 5.22 bpm versus 7.65 bpm for the CNN baseline (Aguiar et al., 2021). A PPG foundation-model study used PPG only at inference but ECG and respiration during pretraining and reported on PPG-DaLiA an across-subject HR MAE 7.78, a within-subject MAE 4.2, and activity-classification MF1 0.36 across-subject and 0.69 within-subject (Geenjaar et al., 5 Jun 2026).

Artifact handling forms a second major cluster. Tiny-PPG cast wrist-PPG artifact detection as sample-wise binary segmentation on 30-second segments and, after pruning, retained 19,726 model parameters (0.15 megabytes) while achieving 87.4% detection accuracy, with the best contrastive configuration reaching 88.4% DICE (Zheng et al., 2023). A learned convolutional sparse-coding denoiser trained on synthetic PulseDB-derived pairs and tested on PPG-DaLiA reduced HR MAE from 11.29 ± 4.39 bpm before denoising to 8.69 ± 3.59 bpm after denoising (Basso et al., 14 Aug 2025). For beat-level inference, a greedy graph-based method reported on PPG-DaLiA Corr 0.91, MAPE 3.8% for stairs and Corr 0.95, MAPE 2.4% for cycling, with HRV metrics such as Mean RR, Mean HR, SDNN, VLF, LF, HF, and Total Power evaluated against ECG-derived references (Huang, 2023).

The dataset has also been repurposed beyond canonical HR regression. In decentralized federated learning, Murmura formulated PPG-DaLiA as an 8-class activity classification problem using time-frequency features and reported accuracies of 63.9% at Dirichlet $\alpha=0.1$, 78.8% at $\alpha=0.5$, and 72.6% at $\alpha=1.0$ (Rangwala et al., 22 Dec 2025). In long-term forecasting, DecompKAN constructed a custom 15-channel physiological benchmark from PPG-DaLiA and achieved best-average MSE 0.568 and MAE 0.424 under the paper’s controlled same-recipe comparison (Mysore, 27 Apr 2026). A reinforcement-learning study mapped PPG-DaLiA physiological streams into a MEWS-based monitoring environment with three agents for heart rate, respiration, and temperature, reporting cumulative rewards rather than conventional clinical metrics (Shaik et al., 2023). The breadth of these uses has made PPG-DaLiA a benchmark ecosystem rather than a single-task corpus.

5. DALIA-1 in HIV immunology

DALIA-1 denotes a phase I trial evaluating a therapeutic vaccine against HIV, based on ex-vivo generated interferon-$\alpha$ dendritic cells loaded with HIV-1 lipo-peptides and activated with lipopolysaccharide (Hejblum et al., 2017). In the cited methodological paper, DALIA-1 is used as a longitudinal flow-cytometry application for a Dirichlet process mixture of multivariate skew $t$-distributions with a sequential posterior approximation strategy.

The analyzed cohort comprised 12 HIV-positive patients with cellular populations quantified at 18 time points, although the reported application focused specifically on week 24 and week 26, immediately after interruption of HAART at week 24. The biological target was the effector CD4$^+$ T-cell population defined as CD45RA$^+$CD27$^-$ among CD3$^+$CD4$\alpha=0.1$0 cells, described as one of the first populations affected during viral rebound (Hejblum et al., 2017).

The methodological contribution lies in borrowing information across repeated measurements without fitting all samples jointly. Using week 24 to inform week 26 through a parametric approximation of the posterior, the sequential model improved the average limited $\alpha=0.1$1-measure from 0.50 for standard NPflow to 0.59, outperforming flowMeans at 0.51. For one highlighted patient, manual gating estimated the effector population at 31.7%, standard NPflow at 7.5%, and sequential NPflow at 38.1% (Hejblum et al., 2017). The work presents DALIA-1 as a case where posterior-informed sequential nonparametric clustering materially improves recovery of a longitudinally perturbed immune-cell subset.

6. DALIA as a declarative agentic layer

In agent systems, DALIA stands for Declarative Agentic Layer for Intelligent Agents. It is proposed as a model-independent architectural layer for MCP-based server ecosystems, intended to reduce hallucinated actions, unexecutable plans, and brittle coordination by explicitly linking goals, capabilities, and execution resources (Rodriguez-Sanchez et al., 24 Jan 2026).

The architecture is organized around four components. First, it introduces a declarative capability representation in which executable operations are described with machine-readable metadata such as functional role, domain, inputs, outputs, preconditions, and postconditions. Second, it defines the Agentic Task Discovery Protocol (ATDP), through which MCP servers expose supported tasks declaratively rather than leaving task inference to prompt-based reasoning. Third, it adds a federated Agent Directory MCP that records agents, roles, domains, and accessible MCP servers. Fourth, it uses these declarations to construct deterministic task graphs grounded exclusively in declared operations (Rodriguez-Sanchez et al., 24 Jan 2026).

The paper’s central design principle is the strict separation of discovery, planning, and execution. Discovery must finish before planning, and planning must finish before execution. The orchestrator then executes the resulting task graph in topological order, propagating intermediate results according to declared input-output dependencies. This architecture is motivated in part by empirical evidence cited from the MAST study, which reported failure rates between 41% and 86% across seven multi-agent frameworks over 1,642 executions (Rodriguez-Sanchez et al., 24 Jan 2026).

DALIA is therefore not a new planning algorithm or a new foundation model. It is an execution-grounding layer that constrains agent behavior to a declared operational space. A plausible implication is that its main contribution is not improving raw reasoning capability, but narrowing the gap between what an agent says it can do and what an underlying server ecosystem can verifiably execute.

7. DALIA as a scalable INLA framework for multivariate spatio-temporal Gaussian processes

In scientific computing, DALIA is a framework for Bayesian inference in multivariate spatio-temporal Gaussian processes, built around integrated nested Laplace approximations (INLA), a sparse inverse-covariance formulation, GPU-accelerated block-dense kernels, and a hierarchical, triple-layer, distributed-memory parallel scheme (Gaedke-Merzhäuser et al., 9 Jul 2025). The framework targets latent Gaussian models whose practical scale exceeds what shared-memory CPU implementations of R-INLA can handle efficiently.

Statistically, the model starts from a latent Gaussian prior with sparse precision $\alpha=0.1$2 and a linear observation model $\alpha=0.1$3. For multivariate processes, DALIA uses a linear model of coregionalization and derives a precision formulation that avoids the enlarged linked-copy representation used in convenient but more expensive interfaces. Computationally, the key step is a permutation that restores a block-tridiagonal with arrowhead structure, making the dominant factorizations, triangular solves, and selected inversion operations amenable to structured GPU kernels (Gaedke-Merzhäuser et al., 9 Jul 2025).

The performance claims are explicitly large-scale. The paper reports weak scaling “surpassing the state-of-the-art by two orders of magnitude on a model whose parameter space is 8× larger,” and strong-scaling speedups of three orders of magnitude on 496 GH200 superchips. In a temporal weak-scaling benchmark, DALIA achieved 124× speedup over R-INLA while solving a model with an 8× larger temporal domain; in a strong-scaling benchmark it reduced iteration time from more than 40 minutes in R-INLA to about 4 minutes on one GPU and then scaled to 496 GPUs with 28.3% parallel efficiency (Gaedke-Merzhäuser et al., 9 Jul 2025).

The framework’s scientific case study modeled PM$\alpha=0.1$4, PM$\alpha=0.1$5, and O$\alpha=0.1$6 over northern Italy for 48 days, using data from the Copernicus Atmosphere Monitoring Service aggregated daily from an original 0.1$\alpha=0.1$7 grid. DALIA then performed spatial downscaling to 0.02$\alpha=0.1$8, a 25× refinement. Reported fixed-effect estimates included an elevation-associated reduction of 0.45 $\alpha=0.1$9 for PM$\alpha=0.5$0, 0.55 $\alpha=0.5$1 for PM$\alpha=0.5$2, and an increase of 1.27 $\alpha=0.5$3 for O$\alpha=0.5$4 per 1 km elevation increase, while estimated pollutant correlations were 0.97 for PM$\alpha=0.5$5–PM$\alpha=0.5$6, -0.61 for PM$\alpha=0.5$7–O$\alpha=0.5$8, and -0.63 for PM$\alpha=0.5$9–O$\alpha=1.0$0 (Gaedke-Merzhäuser et al., 9 Jul 2025). This use of DALIA is therefore both methodological and application-driven: it is a statistical-HPC stack whose significance depends on making previously impractical multivariate spatio-temporal inference routine at supercomputing scale.