Open DAC 2025: A Multidomain Platform

• ODAC25 is a multidomain open research label that defines distinct technologies in ion-trap hardware, neural audio coding, and carbon capture datasets.
• Each implementation emphasizes modular design and performance engineering, ensuring low noise in hardware, reproducible ML results, and corrected DFT datasets.
• The initiative promotes reproducibility and scalability via open-source frameworks, enabling cross-disciplinary benchmarks and collaborative advancements.

Searching arXiv for papers and exact ODAC25-related sources. Using the arXiv search tool to confirm ODAC25-related papers and disambiguate the term across domains. Open DAC 2025 (ODAC25) is a designation that appears in multiple 2025-era research contexts rather than a single universally fixed artifact. In the current arXiv record, it denotes an open-hardware, low-noise digital-to-analog control platform for ion-trap electrodes grounded on the Vanguard DAC System-on-Module; it also appears as an open benchmarking and deployment context for a JAX implementation of the Descript Audio Codec; and it names a large-scale dataset for sorbent discovery in direct air capture from humid air (Peaks et al., 1 May 2026, Braun, 2024, Sriram et al., 5 Aug 2025). Across these usages, ODAC25 is associated with openness, modularity, and scale, but the technical meaning depends entirely on domain.

1. Terminology and domain-specific usages

The acronym “DAC” is overloaded across several research areas, and ODAC25 inherits that ambiguity. In ion-trap control, DAC denotes digital-to-analog conversion hardware; in machine learning for audio, DAC denotes the Descript Audio Codec; in carbon capture, DAC denotes direct air capture. The label ODAC25 is therefore best understood as a domain-qualified term.

Domain	ODAC25 usage	Core technical object
Ion-trap control	Open DAC 2025 platform	32-channel, 16-bit DAC SoM based on DAC81416 and Spartan-7
Neural audio coding	Open DAC initiative context	JAX/Flax implementation of the Descript Audio Codec
Carbon capture	Open DAC 2025 dataset	Nearly 70 million DFT single-point calculations over ~15,000 MOFs

This multiplicity is not merely terminological. Each usage organizes a different methodological stack: mixed-signal electronics and FPGA gateware in trapped-ion systems, differentiable ML infrastructure in neural audio codecs, and DFT plus MLIP pipelines in materials discovery. A plausible implication is that ODAC25 functions more as a 2025 open-research label than as a single standardized platform name.

2. ODAC25 as an open-hardware DAC platform for ion-trap electrode control

In the ion-trap context, ODAC25 can be grounded on, or directly adopt, the Vanguard DAC System-on-Module described in “Low-cost Ultra-low Noise DAC System-on-Module for Scalable Ion-Trap Electrode Control” (Peaks et al., 1 May 2026). The architecture is centered on two Texas Instruments DAC81416 devices and an AMD Xilinx Spartan-7 FPGA (XC7S6-1FTGB196C), yielding 32 independent analog channels per module. The DAC81416 is a 16-channel, 16-bit DAC with internal 2.5 V precision reference; the SoM uses two such devices, identified in firmware as DAC_0 and DAC_1. The FPGA handles UART-to-SPI data movement, configuration, triggering, and synchronization, and is clocked from an on-board 50 MHz MEMS oscillator, while the minimum viable gateware operates internal logic at 10 MHz.

The module’s present power architecture is set to ±10 V rails, although the DAC81416 supports configurable bipolar ranges up to ±20 V. At ±10 V, the 16-bit LSB is approximately 305.176 µV per code step. The internal reference accuracy is specified as ±2.5 mV max with typical drift of about 5 ppm/°C. Serial communication to the DACs uses 4-wire SPI up to 50 MHz; in the prototype, the SPI cores run at 10 MHz. Host control is via a USB-to-UART cable connected to an FPGA UART core, with a Python control script on the host. An external hardware trigger is provided through a single SMA input wired to FPGA GPIO.

The output stage emphasizes low noise and mixed-signal isolation. The SoM uses two-stage linear regulation across digital and analog domains and explicitly avoids on-board switching regulators. Front-end regulators include L7815, MC7915ACD2TG, L7806, and L7805, followed by secondary low-noise LDOs including TI TPS7A8801, ADI LT3042, and ADI LT3032-12. The PCB is an 8-layer design with physically separated analog and digital sides, extensive ground stitching, and EMI wall vias. The 37-pin micro-D connector carries analog outputs; in Rev A there are no explicit ground pins on the output connector, so the single-board ground reference is used.

Performance characterization indicates suitability for ion-trap physics experiments and quantum computing applications. The DAC81416 datasheet settling time is about 12 µs typical, and the nominal slew rate is about 4 V/µs. Each channel includes a single-pole RC low-pass filter with measured cutoff at 48.22 kHz. For that pole,

$f_c = \frac{1}{2\pi RC},$

and the effective time constant is approximately 3.3 µs, implying about 16.5 µs to 1% settling. On the SoM, measured full-scale step slew with filters enabled was an average of 1.76 V/µs for a 20 V step from −10 V to +10 V. Measured near-zero codes gave an average LSB increment of 308(32) µV, consistent with the expected 305.176 µV. Device-to-device static offset between DAC_0 and DAC_1 was 1.07(3) mV and can be calibrated out.

The design is explicitly targeted at scalable electrode control. The DAC81416 supports asynchronous mode, synchronous mode, streaming mode, and SPI daisy-chaining. In the present build, asynchronous register writes update outputs immediately, but synchronous mode would permit write-then-trigger simultaneous updates across channels, which is directly relevant to time-critical ion transport. The paper also recommends shared clocks, shared triggers, star-grounded backplanes, shielded enclosures, and external precision references when scaling to many modules. This suggests that ODAC25, in this usage, is not merely a board-level design but a modular control-plane concept for multi-hundred-channel trapped-ion systems.

3. ODAC25 in neural audio codec research

In the audio domain, ODAC25 refers to an open initiative built around “DAC-JAX: A JAX Implementation of the Descript Audio Codec” (Braun, 2024). Here DAC denotes the Descript Audio Codec, described as a high-fidelity neural audio codec built around an autoencoder with residual vector quantization and a convolutional discriminator under adversarial training. The JAX implementation re-creates the PyTorch DAC in the JAX ecosystem using Flax for models, Optax for optimizers, Orbax for checkpointing, AUX for audio and STFT utilities, and CLU for metrics and training loops.

The model details reported in the paper are precise. For the 44.1 kHz “8 kbps” large model, the encoder strides are $2\times 4\times 8\times 8 = 512$, which gives a receptive field of 512 samples and a token rate of approximately $44100/512 \approx 86$ Hz per codebook stream. There are $M=9$ codebooks, each producing 10-bit tokens, so each codebook has $K=1024$ entries, and each token maps to an 8-dimensional embedding. The effective bitrate per channel is approximately

$$\left(\frac{44100}{512}\right)\times 9 \times 10 \approx 7752\ \text{bps},$$

which is below 8 kbps.

A central technical result is full weight compatibility and implementation equivalence. The codebase can reuse weights from the original PyTorch DAC, and the authors confirm that for the same input, DAC-JAX produces equivalent token sequences and decoded audio. Chunked compression and decompression for long audio are also re-implemented, with window size rounded to a multiple of the receptive field and hop size derived from encoder and decoder rates to match PyTorch scheduling semantics. DAC-JAX requires explicit instantiation with padding=False for chunked operations and a separate padding=True instance for training, reflecting differences imposed by Flax’s functional style.

The implementation is also positioned as a systems comparison between ML ecosystems. Training and fine-tuning scripts support device parallelism via pmap; augmentations run on-device to avoid CPU bottlenecks; tf.data provides JAX-native dataset loading; jaxloudnorm replaces pyloudnorm; and Julius resampling is re-implemented in JAX. Benchmarks show a hardware-dependent crossover. On an Nvidia RTX 2080, DAC-JAX is faster than PyTorch DAC for both compression and decompression across all tested chunk sizes. On an Nvidia L40, JAX is faster at small hops such as 8.2 ms and 646.8 ms, but slower for larger hops such as 4640.6 ms and 19640.7 ms. The paper attributes this pattern plausibly to XLA compilation and launch overhead, kernel granularity, memory bandwidth, and host-device synchronization behavior.

Within this usage, ODAC25 denotes an open, weight-compatible JAX-native baseline for neural audio coding research. Its importance is methodological rather than taxonomic: it provides a reproducible codec implementation, chunked inference primitives, performance scripts, and training utilities that make PyTorch-versus-JAX comparisons concrete at the level of token equivalence, throughput, and deployment workflow.

4. ODAC25 as a dataset for sorbent discovery in direct air capture

The most explicit use of the name as a standalone title appears in “The Open DAC 2025 Dataset for Sorbent Discovery in Direct Air Capture” (Sriram et al., 5 Aug 2025). In this context, ODAC25 is a dataset for identifying sorbent materials for direct air capture from humid air. It comprises nearly 70 million DFT single-point calculations for CO$_2$, H$_2$O, N$_2$, and O$_2$ adsorption in approximately 15,000 metal-organic frameworks, and is described as a substantial expansion over ODAC23.

Its scope is defined by both scale and configurational diversity. ODAC25 includes ODAC23-style placements generated from Monte Carlo sampling with classical force fields, followed by full DFT relaxation with all atoms allowed to move. It adds high-energy GCMC-derived placements by taking non-relaxed snapshots from GCMC trajectories at 300 K and pressures of 5, 10, 20, and 50 kPa for pure CO$2\times 4\times 8\times 8 = 512$0, pure H$2\times 4\times 8\times 8 = 512$1O, and mixtures with gas-phase molar ratios of 1:1, 1:5, 1:10, and 1:20. Random snapshots across 500,000 steps per simulation yielded over 2.7 million single-point DFT calculations. ODAC25 also expands chemistry via linker functionalization, open-metal-site functionalization with diamines, reactive CO$2\times 4\times 8\times 8 = 512$2 placements meant to capture ammonium–carbamate chains and carbamic acid mechanisms, and a synthetic MOF split derived from an autoregressive transformer and assembled with Pormake.

The dataset also standardizes and upgrades electronic-structure methodology. All DFT calculations use VASP 6.3 with PBE and D3 dispersion, VASP 5.4 PBE pseudopotentials, a plane-wave cutoff of 600 eV, and electronic convergence of $2\times 4\times 8\times 8 = 512$3 eV. Gaussian smearing with $2\times 4\times 8\times 8 = 512$4 eV, spin polarization, and symmetry-off settings are applied. Structural relaxations use conjugate gradient; bare MOFs are relaxed with ISIF=3 and MOF+adsorbate systems with ISIF=2. ODAC25 further corrects a systematic ODAC23 issue arising from uniform $2\times 4\times 8\times 8 = 512$5 k-point grids by using per-trajectory energy offsets relative to a higher-density reference grid. The paper states that about 7% of systems had greater than 0.2 eV total-energy error under ODAC23 settings, and that the two-point correction reduced energy errors to about 0.01 eV at under 1% of the cost of fully rerunning all frames.

A second methodological correction concerns flexible frameworks and adsorption-energy referencing. For each converged MOF+adsorbate relaxation, the empty framework is re-relaxed, sometimes yielding multiple conformers for a single MOF. Adsorption energies are then referenced to the lowest-energy empty MOF discovered across all such re-relaxations. The paper reports median adsorption-energy shifts greater than 0.1 eV toward less favorable values across splits, removing physically inconsistent references. This is important because ODAC25 is intended not only as a repository of DFT points but as a physically corrected dataset for screening and ML supervision.

The release also includes machine-learned interatomic potentials. The principal models are eSEN and UMA, with comparisons to EquiformerV2-Large and MACE variants. Training targets are total energies and forces rather than adsorption energies directly, with adsorption quantities computed post hoc. The reported loss is

$2\times 4\times 8\times 8 = 512$6

For the filtered ODAC25 MOF+adsorbate test set, eSEN (Full) achieves EMAE-Tot $2\times 4\times 8\times 8 = 512$7 eV, adsorption EMAE $2\times 4\times 8\times 8 = 512$8 eV, and force MAE $2\times 4\times 8\times 8 = 512$9, while eSEN (Filtered) gives adsorption EMAE $44100/512 \approx 86$0 eV and force MAE $44100/512 \approx 86$1. By contrast, the ODAC23 EquiformerV2-Large baseline is reported with adsorption EMAE $44100/512 \approx 86$2 eV. Henry’s law coefficients are computed via Widom insertion:

$44100/512 \approx 86$3

and compared against experiments for UiO-66, HKUST-1, and MOF-5. The paper reports that UMA-Medium 1.1 performs best on CO$44100/512 \approx 86$4 and eSEN (Filtered) performs best on N$44100/512 \approx 86$5 in those Henry-coefficient comparisons.

In this sense, ODAC25 is a data-and-model infrastructure for humid-air DAC screening. Its technical significance lies in combining larger chemical coverage, non-minimum-energy states, corrected adsorption-energy references, and MLIP training targets based on total energies and forces.

5. Shared design motifs across the three ODAC25 usages

Despite the disparate subject matter, the three ODAC25 usages show a striking convergence in research style. Each combines openness with a scaling strategy. The ion-trap ODAC25 design emphasizes in-stock components, free programming tools, an open repository, and architectural provisions for daisy-chain SPI, shared clocks, shared triggers, and star-grounded backplanes. DAC-JAX is explicitly open-source under an MIT license, reproduces PyTorch behavior in JAX, and provides chunked inference plus device-parallel training primitives. The direct-air-capture ODAC25 release combines a large dataset with accessible code resources through FAIRchem and a Widom-insertion implementation built on ASE calculators.

A second commonality is the coupling of fidelity constraints to performance engineering. In the ion-trap system, linear regulation, RC filtering, and analog-digital isolation are used to preserve low-noise DC control while still supporting transport-relevant update speeds. In DAC-JAX, equivalence of token sequences and decoded audio is preserved while chunk size and JIT compilation are used to manage memory footprint and throughput. In the direct-air-capture dataset, k-point corrections, re-relaxations, and total-energy MLIP targets are introduced to improve physical consistency without rerunning the entire corpus at full cost.

A third commonality is modularity. The trapped-ion platform is organized as a 32-channel SoM intended for multi-module expansion. DAC-JAX separates chunked inference instances from training instances and exposes per-chunk JIT-compiled functions. ODAC25 for direct air capture is released in “full” and “filtered” splits, and its synthetic, functionalized, and high-energy GCMC-derived subsets can be used selectively. This suggests that ODAC25, across fields, is associated less with monolithic systems than with open modular stacks whose interfaces remain visible to the researcher.

6. Misconceptions, boundary cases, and relation to earlier DAC literature

A common misconception is that ODAC25 refers to a single benchmark or hardware platform. The arXiv record instead supports a tripartite interpretation: an ion-trap control module lineage, an audio-codec implementation lineage, and a direct-air-capture dataset lineage. Another potential misconception is that these usages are versions of one another. They are not: they share a naming pattern and an open-research orientation, but they belong to different technical communities and use incompatible meanings of “DAC.”

This ambiguity is sharpened by earlier literature in which DAC stands for Distributed Arithmetic Coding. Fang’s work on DAC codeword spectrum defines path spectrum, time spectrum, and an expansion factor for Slepian–Wolf decoding, and proves that when symbols $44100/512 \approx 86$6 and $44100/512 \approx 86$7 are mapped onto intervals $44100/512 \approx 86$8 and $44100/512 \approx 86$9 with $M=9$0, the expansion factor converges to $M=9$1 as decoding proceeds (Fang, 2010). That result is mathematically important for distributed source coding, but it is terminologically separate from ODAC25 in ion-trap electronics, neural audio codecs, and direct-air-capture materials datasets.

The main boundary condition, therefore, is semantic. Any technical discussion of ODAC25 must specify the domain at the outset. In trapped-ion control it denotes a low-noise, scalable DAC SoM reference design; in audio ML it denotes an open JAX realization of the Descript Audio Codec suitable for reproducible benchmarking; and in carbon-capture materials science it denotes a large DFT-and-MLIP dataset for humid-air sorbent discovery. Without that qualification, the acronym is inherently polysemous.