Accelerated AI Algorithms for Data-Driven Discovery

What Accelerated AI Algorithms for Data-Driven Discovery does

A3D3 (Accelerated AI Algorithms for Data-Driven Discovery) is a federally funded, multi-institutional research institute established in 2021 under the National Science Foundation's Harnessing the Data Revolution (HDR) program with approximately $15 million in initial funding (award PHY-2117997). Its mission is to develop real-time AI/ML algorithms deployed on heterogeneous computing platforms — CPUs, GPUs, FPGAs, and ASICs — to accelerate scientific discovery in three domains: high energy physics (notably processing LHC collision data at 40 MHz), multi-messenger astrophysics (gravitational-wave detection via LIGO, neutrino detection via IceCube, transient classification via ZTF), and systems neuroscience (real-time neural state detection from high-density electrophysiology). The institute's core technical contribution is hardware-algorithm co-design, embodied in a suite of open-source tools (hls4ml, PyLog, SONIC, ScaleHLS, QONNX, ML4GW, NMMA, TorchSparse++) that translate ML models into FPGA/ASIC firmware with documented 50-300x speedups over CPU baselines.

The institute operates across 14+ academic institutions led by University of Washington (Director Shih-Chieh Hsu) and MIT (Deputy Director Philip Harris), with principal investigators at Caltech, Duke, UCSD, UIUC, University of Minnesota, Purdue, UW-Madison, Georgia Tech, and affiliate institutions including National Yang Ming Chiao Tung University (Taiwan), University of Pennsylvania, Lawrence Berkeley National Laboratory, and Westmont College. Governance is distributed across collaborating PIs; there is no single legal entity, equity ownership, or venture backing. All software, tutorials, and educational materials are distributed free of charge under open-source licenses via GitHub, with adoption driven through scientific collaborations, conference presentations, monthly seminars, mentoring programs, and an annual All-Hands Meeting.

A3D3 has no commercial revenue. Its sole revenue stream is NSF government grant funding, supplemented by individual PI awards (DOE Early Career, NSF CAREER, Sloan Fellowship) at the participating universities. Pricing is non-applicable as the institute does not sell products or services. Its customer base — in a mission-impact sense rather than commercial sense — consists of major science experiments including ATLAS, CMS, LIGO/Virgo/KAGRA, IceCube, ZTF, and DUNE, where A3D3-developed AI tools are integrated into production data processing pipelines.

Name

Accelerated AI Algorithms for Data-Driven Discovery

Legal name

Accelerated AI Algorithms for Data-Driven Discovery (A3D3)

Website

https://a3d3.ai

Company type

Private

Founded year

2021

Operating status

Operating

Headcount range

1–10 employees

Short description

A3D3 is an NSF-funded, multi-institutional research institute developing real-time AI/ML algorithms and hardware-algorithm co-design tools for scientific discovery in high energy physics, multi-messenger astrophysics, and systems neuroscience, deployed at major science experiments including CERN, LIGO, IceCube, ZTF, and DUNE.

Ownership category

akta.pro rank

Product category

Scientific Machine Learning Software

NAICS

Scientific Research and Development Services (5417), Research and Development in the Physical, Engineering, and Life Sciences (except Nanotechnology and Biotechnology) (541715), Computer Systems Design and Related Services (54151), Computer Systems Design and Related Services (5415)

SIC

Services-Computer Programming Services (7371), Services-Computer Programming, Data Processing, Etc. (7370), Services-Computer Integrated Systems Design (7373)

akta.pro primary industry

AI Compiler, Runtime & Kernel Optimization Software (CUDA/ROCm/XLA, graph compilers) (HDAAAAAI)

akta.pro secondary industries

AI Server Systems & HGX/Accelerator Platforms (HDAAAAAB), GPU-Accelerated & AI Training/Inference Servers (HDACABAG), Programmable Logic (FPGAs/CPLDs) (HDAHAJAG), Synthetic Data & Data Augmentation for Foundation Models (HDAAACAK), On-Device Inference Runtimes & SDKs (mobile/embedded) (HDAAAJAB)

Keywords

• Scientific machine learning
• Real-time AI inference
• FPGA hardware acceleration
• Particle physics algorithms
• Multi-messenger astrophysics

Headquarters

HQ city

Seattle

HQ country

United States

HQ region

North America

Offices14 records

Markets served

GTM type

B2B

Offering type

Software

Cost components

Personnel, Technology or R&D, Operations, Infrastructure, Others

Revenue model

1. NSF HDR Program Funding: A3D3 is funded by the National Science Foundation (NSF) under the Harnessing the Data Revolution (HDR) program. The institute represents a multi-institutional collaborative entity with NSF grant PHY-2117997 providing the primary funding for research operations, personnel, equipment, and outreach activities.

Go-to-market motion2 records

Distribution channels4 records

Marketing channels13 records

Core offering

A3D3 develops and disseminates accelerated AI/ML algorithms, software frameworks, and hardware-algorithm co-design tools that enable real-time processing and analysis of massive scientific datasets. The institute builds open-source libraries spanning FPGA-accelerated inference, sparse neural network training, probabilistic programming, and graph-based reconstruction for high-energy physics, multi-messenger astrophysics, and neuroscience. Deliverables are openly released and integrated into leading scientific experiments' data processing pipelines.

Differentiator

Problem solved

Functional benefit

Products and services

• hls4ml High-level synthesis framework that converts trained neural networks into hardware descriptions deployable on FPGAs, enabling nanosecond-scale inference for real-time triggers and event filtering in high-energy physics detectors.
• ML4GW Software toolkit applying ML to gravitational-wave signal detection, classification, and parameter estimation for observatories such as LIGO.
• NMMA Code package for multi-messenger astrophysics analysis that jointly models electromagnetic and gravitational-wave signatures of compact binary mergers.
• SONIC Compiler and runtime system that accelerates sparse neural network inference by exploiting sparsity in weights and activations to reduce compute and memory requirements.
• TorchSparse GPU-accelerated sparse tensor training library that enables efficient training of neural networks on sparse 3D point cloud data such as detector hits.
• PyLog Probabilistic programming library that combines physics simulation with Bayesian inference for scientific modeling in physics and astrophysics.
• ScaleHLS Compiler framework that scales high-level synthesis flows to large ML models, enabling efficient translation to FPGA and ASIC hardware implementations.
• HIDA Data augmentation framework that injects realistic detector effects into training data to improve the robustness and generalization of physics ML models.
• QONNX Quantized neural network interchange format and supporting tooling that enables portable deployment of quantized models across hardware accelerators.
• aframe Production deep-learning pipeline for real-time gravitational-wave detection and rapid sky localization built on the aframe architecture.
• GWAK ML pipeline that performs anomaly detection and classification on gravitational-wave data streams to identify rare or novel signal classes.

Quantifiable outcome

• 4% higher mIoU and 10+% higher PQ than GravNet baseline for Hadron Calorimeter particle reconstruction using TorchSparse sparse CNN
• +5 more outcomes

Named customers7 records

Segments4 records

Ideal customer profiles2 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

Integration3 records

AI capability12 records

Feature15 records

Partnerships

14 partnerships are on record, tiered strategic, core and minor.

• AMD (Xilinx)strategicTechnology or IntegrationAMD/Xilinx is listed as an industry partner. A3D3 tools including hls4ml, ScaleHLS, and PyLog support AMD/Xilinx FPGA platforms (Vivado HLS, Vitis HLS) and the company provides hardware and technical collaboration for ML deployment research.
• ATLAS CollaborationcoreStrategic or Co-development PartnerThe ATLAS experiment at CERN is a primary science collaboration partner. A3D3 researchers develop and deploy AI tools for ATLAS hardware triggers using FPGAs, on-detector ASIC processing, and anomaly detection methods for the LHC detector.
• CMS CollaborationcoreStrategic or Co-development PartnerThe CMS experiment at CERN is a primary science collaboration partner. A3D3 develops and deploys AI solutions for CMS including SONIC (Services for Optimized Network Inference on Coprocessors), hls4ml-based trigger algorithms, and ML4GPU-accelerated data reconstruction.
• LIGO Scientific CollaborationcoreStrategic or Co-development PartnerLIGO (Laser Interferometer Gravitational-wave Observatory) at Caltech/MIT is a primary science collaboration partner. A3D3 developed aframe (first end-to-end ML-based real-time search for binary black holes) and GWAK (Gravitational-Wave Anomalous Knowledge) for LIGO data.
• IceCube Neutrino ObservatorycoreStrategic or Co-development PartnerIceCube at the South Pole is a science collaboration partner. A3D3 accelerates ML inputs for IceCube by improving calibration and feature extraction for individual channels, reducing power requirements and time to result.
• Zwicky Transient Facility (ZTF)coreStrategic or Co-development PartnerZTF at Palomar Observatory/Caltech is a science collaboration partner. A3D3 developed neural network and XGBoost algorithms for ZTF source classification into a public catalog, and trains summer school participants in ZTF data analysis.
• Deep Underground Neutrino Experiment (DUNE)coreStrategic or Co-development PartnerDUNE is a primary science collaboration partner. A3D3 researchers develop ML-based supernova neutrino direction reconstruction using 1D CNNs and graph neural networks for DUNE's liquid argon time projection chamber.
• FastMachineLearning ConsortiumstrategicStrategic or Co-development PartnerFastMachineLearning is a science collaboration partner focused on ML for physics. A3D3 co-develops hls4ml, SONIC, SuperSONIC, and QONNX within the FastMachineLearning ecosystem, and jointly organizes the FastMachineLearning Workshop.
• IRIS-HEP (Institute for Research and Innovation in Software for High Energy Physics)strategicStrategic or Co-development PartnerIRIS-HEP is an NSF-funded software institute for HEP. A3D3 collaborates with IRIS-HEP on software infrastructure, computing frameworks, and training for particle physics.
• AI Institute for Artificial Intelligence and Fundamental Interactions (IAIFI)strategicStrategic or Co-development PartnerIAIFI is another NSF HDR institute focused on AI for fundamental physics. A3D3 collaborates with IAIFI on shared research challenges, workshops, and transdisciplinary initiatives within the HDR ecosystem.
• ID4 (Institute for Data Driven Dynamical Design)minorStrategic or Co-development PartnerID4 is one of five NSF HDR institutes. A3D3 collaborates with ID4 on HDR ecosystem activities including annual conferences, workshops, and shared education resources.
• Imageomics InstituteminorStrategic or Co-development PartnerImageomics is one of five NSF HDR institutes. A3D3 collaborates with Imageomics on HDR ecosystem activities including annual conferences, education repository development, and workshops.
• iHARP (Institute for Harnessing Data and Model Revolution in the Polar Regions)minorStrategic or Co-development PartneriHARP is one of five NSF HDR institutes. A3D3 collaborates with iHARP on HDR ecosystem activities including annual conferences and shared workshops.
• I-GUIDE (Institute for Geospatial Understanding through an Integrative Discovery Environment)minorStrategic or Co-development PartnerI-GUIDE is one of five NSF HDR institutes. A3D3 collaborates with I-GUIDE on HDR ecosystem activities including annual conferences, education repository development, and transdisciplinary workshops.

Scale indicators5 records

Recent moves8 records

Expansion highlights6 records

Direct peers

• FastMachineLearning Consortium: Open consortium co-developing hls4ml, SONIC, and QONNX with A3D3; focuses on ultra-low-latency ML for physics experiments. Direct peer in tooling and community.
• Imageomics Institute: One of five NSF HDR institutes, focused on AI for biological image data. Sister institute under the same NSF funding mechanism and collaborates with A3D3 on HDR ecosystem activities.
• IAIFI (AI Institute for Artificial Intelligence and Fundamental Interactions): Sibling NSF HDR institute focused on AI for fundamental physics; collaborates with A3D3 on workshops and shared research challenges. Directly comparable in funding model, scope, and target scientific community.
• IRIS-HEP (Institute for Research and Innovation in Software for High Energy Physics): NSF-funded software institute explicitly co-developed with A3D3; both focus on software infrastructure for HEP, share trainees, and collaborate on compute and training programs. Highly comparable in mission and structure.

Emerging players

• Cerebras Systems: Builds AI accelerator hardware (wafer-scale CS systems) used in scientific computing, including HEP collaborations. Emerging peer in the heterogeneous AI compute space relevant to A3D3's hardware-algorithm co-design mission.

Broad incumbents

• AMD / Xilinx (Vitis, Vivado HLS): Strategic industry partner of A3D3; hls4ml supports Xilinx HLS backends. AMD/Xilinx provides the FPGA fabric that A3D3's tools target, making it a closer commercial peer in the HLS-to-FPGA workflow.
• NVIDIA (HLS / FPGA / Edge AI efforts): Commercial vendor of CUDA, TensorRT, Triton, and adjacent AI infrastructure; A3D3 explicitly integrates with NVIDIA Triton. A broad incumbent providing overlapping AI compute capabilities though focused on general-purpose GPU rather than FPGA/ASIC scientific pipelines.
• CERN openlab: CERN's public-private R&D partnership program engages industry and academia on compute, AI, and data challenges for the LHC. A3D3 collaborates directly with ATLAS and CMS at CERN, making openlab a broad institutional peer.
• Synopsys (HLS Toolchain): Commercial EDA vendor with HLS tools (Synphony HLS, Catapult) targeted at FPGA and ASIC design. Comparable in offering HLS compiler infrastructure overlapping with A3D3's ScaleHLS, HIDA, and hls4ml.

Regional players

• Allen Institute for Brain Science: Non-profit research organization producing large-scale neuroscience datasets and AI tools for brain mapping. Comparable to A3D3's systems neuroscience thrust in being a research-driven AI-for-science entity, though complementary rather than directly competitive.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat6 records

Key risks5 records

Key highlights6 records

Customer concentration

• Xx.com/a3d3institute
• GitHubgithub.com/a3d3-institute
• YouTubeyoutube.com/channel/UCozE47hFawTUb4lUhYlQNIQ
• LinkedInlinkedin.com/company/a3d3

Revenue estimate

Valuation estimate

Number of profiles

Profiles2 records

Funding overview

Funding rounds

Investors

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Accelerated AI Algorithms for Data-Driven Discovery