Scalable Platform for Enterprise Engineering and Deployment towards Mathematics for the Discovery of Algorithms and Architectures (SPEED DIAL)

Scalable Platform for Enterprise Engineering and Deployment towards Mathematics for the Discovery of Algorithms and Architectures (SPEED DIAL) is sponsored by DARPA. Develop and demonstrate a robust framework and tool suite building on automated algorithm discovery tools that will enable the seamless integration of discovered algorithms with the scientific and engineering workflow.

SPEED DIAL | STTR | DARPA Department of War organization.

Scalable Platform for Enterprise Engineering and Deployment towards Mathematics for the Discovery of Algorithms and Architectures Partner with Us to Accelerate DARPA Innovation OUSD (R&E) critical technology area(s) : Advanced Computing and Software, Applied Artificial Intelligence, Human-Machine Interfaces, Integrated Network Systems-of-Systems Objective: Develop and demonstrate a robust framework and tool suite building on automated algorithm discovery tools that will enable the seamless integration of discovered algorithms with the scientific and engineering workflow.

This framework will provide domain experts with on-demand tools to discover new, bespoke algorithms, bridging the gap between algorithmic discovery and practical application and empowering engineers and scientists to leverage AI-driven algorithm design as a standard part of their process.

Description: The DARPA DIAL [1] program has successfully demonstrated that AI can autonomously discover novel, high-performance algorithms for a range of scientific and engineering problems.

Recent efforts have shown that models such as Transformers can rediscover fundamental algorithms like the Kalman Filter, Genetic Programming can rediscover Wavelets, Accelerated Monte Carlo Tree Search can rediscover optimization algorithms, and AI-driven methods can produce optimal meta-solvers for complex physics simulations.

However, such advanced discovery engines are currently siloed in research environments; they are not yet deployed for scientists to discover new algorithms before they execute their standard scientific process. This project will address this gap by creating a framework to transition both previously discovered algorithms and the algorithm discovery engines themselves from curiosities to commodities [2-5].

The goal is to create a symbiotic relationship between algorithm developers and domain experts (e.g., engineers, physicists), enabling them to collaboratively discover, build, refine, and deploy algorithms for real-world applications.

We will foster a partnership between universities (with deep expertise in algorithmic discovery) and companies (representing the US industrial base) to embed these tools directly into standard workflows, allowing engineers to generate optimal algorithms on-the-fly for their specific constraints. This topic is soliciting Direct to Phase II (DP2) proposals only.

Phase I Feasibility (for Direct to Phase II): Proposers must provide written evidence (can include technical reports, peer-reviewed publications, etc.) of the following: Discovered Novel Algorithms: Successfully developed and employed methods (e.g., Transformers, Genetic Programming, Reinforcement Learning) to discover novel algorithms that outperform existing state-of-the-art methods in domains such as time-series analysis, data compression, and solving complex partial differential equations.

Demonstrated Performance Gains: Quantitatively shown the benefits of these discovered algorithms, achieving significant speedups and performance improvements. For example, previous research has demonstrated that AI-discovered wavelets can outperform existing ones by a factor of 1000x, and novel AI-generated meta-solvers have shown a 6. 35x reduction in iterations for complex acoustic and sonar problem sets.

Achieved Human Interpretability: The discovered algorithms are not "black boxes." They are composed of transparent building blocks that are understandable to domain experts, a crucial factor for adoption, verification, and certification. The Phase II effort will focus on building a comprehensive framework to transition algorithm discovery and deployment into mainstream scientific and engineering practice.

The work will be divided into the following key tasks: Develop a Unified Discovery and Deployment Platform: Create a software platform that serves as a seamless interface between AI-driven algorithm discovery engines and domain-specific simulation/modeling environments. This platform will allow engineers to define their problem spaces, boundary conditions, and hardware constraints to initiate the discovery process.

Pervasive Algorithm Discovery Engine: Embed active discovery mechanisms (e.g., AlphaEvolve paradigms, Transformers, Genetic Programming) directly into the platform. This will enable industrial partners to move beyond pre-existing solutions and autonomously discover new, bespoke algorithms optimized specifically for their proprietary datasets and novel engineering challenges.

Establish a "Discovered Algorithm" Library: Curate an evolving, version-controlled library of the most promising algorithms. This library will be organized by problem class (e.g., linear algebra, optimization, signal processing) and will include performance benchmarks and usage guidelines to act as a "warm start" for new problems.

Create "In-Context" Integration Tools: Develop tools that allow engineers to seamlessly integrate both newly discovered and library-retrieved algorithms into their existing workflows and software (e.g., MATLAB, Simulink, COMSOL, custom C++ environments). Crucially, "in-context" in this framework means combining the problem description and the algorithmic solution together as a unified pair.

By encoding the physical problem constraints directly with the mathematical solution structure (inspired by methodologies like GRAFT-ATHENA [6]), the system achieves vastly superior, context-aware retrieval. Once retrieved, these tools will utilize APIs, wrappers, and auto-compilation to deploy the algorithm and abstract away the underlying code complexity.

Demonstrate in Real-World Scenarios: In partnership with industry and DoW stakeholders, performers will demonstrate the framework on at least two distinct, DoW-relevant problems. Potential application areas include, but are not limited to, hypersonic vehicle design, submarine acoustic signature analysis, or digital twin modeling for advanced systems design and manufacture.

Success will be measured by the ability of non-algorithm experts to discover and deploy a novel algorithm that yields quantifiable improvements in efficiency and time-to-solution for the right problem. Month 4: Architecture & Initial Library Deliverable: Prototype of the active, ambient SPEED DIAL Discovery Engine.

Criteria: Successful integration of background discovery algorithms (including Monte Carlo Tree Search and Genetic Programming) running on local/cloud nodes that continuously generate and evaluate novel mathematical variations without interrupting the primary engineering workflow. Month 8: Pervasive Engine Integration Deliverable: System Architecture Document and Initial Algorithm Library.

Criteria: Platform interface mockups completed; library populated with at least 5 foundational algorithms (e.g., optimized wavelets, meta-solvers) ready for deployment. Month 12: In-Context Retrieval & Workflow Integration Tools Deliverable: API Wrappers, Auto-compilation Toolchain, and Problem-Solution Pair Encoding Interface.

Criteria: Successful "In-Context" deployment of a library-derived algorithm into a standard commercial environment (e.g., MATLAB) without manual code rewriting. Demonstration of joint problem-solution pairing to retrieve context-aware algorithms. Month 14: Interim DoW Demonstration Deliverable: DoW Use Case 1 Demonstration Report.

Criteria: Non-algorithm experts successfully utilize SPEED DIAL on a DoW-relevant problem (e.g., hypersonic design), demonstrating >10% improvement in computational efficiency. Month 18: Closed-Loop Pipeline Release Deliverable: Beta Release of the integrated SPEED DIAL Platform.

Criteria: End-to-end user flow validated: user defines a problem, and the system ambiently discovers a new algorithm, compiles it, and suggests a deployable API wrapper update while the user continues to work. Associated Tasks: 1, 2, 4 Month 24: Final DoW Demo & Transition Deliverable: DoW Use Case 2 Demonstration, Final Software Release, and Phase III Transition Plan.

Criteria: Deployment of a second major DoW application demonstrating pervasive, bespoke discovery yielding scalable performance gains across the industrial base. Phase III dual use applications The framework developed in this STTR will have broad applicability across both the DoD and commercial sectors.

DoD/Military Applications: Hypersonics: Accelerated design and testing of hypersonic vehicles through faster and more accurate CFD simulations. Sonar and Radar: Improved signal processing for target detection and classification. Logistics and Supply Chain: Optimization of complex logistics networks for contested environments.

Digital Engineering: Rapid development and validation of digital twins for a wide range of military systems. Advanced Manufacturing: Optimization of manufacturing processes for improved efficiency and quality. Financial Modeling: Discovery of novel algorithms for risk analysis and high-frequency trading.

Drug Discovery: Accelerated simulation of molecular dynamics for the development of new pharmaceuticals. Semiconductor Design: Optimization of chip layouts and manufacturing processes. [1] DARPA DIAL Program Announcement (DARPA-PA-23-03-12); https://sam.

gov/workspace/contract/opp/8e9428a0783e483f954d93555dd8cc77/view [2] Discovering Algorithms with Computational Language Processing; https://arxiv. org/pdf/2507. 03190 [3] Learning to Discover Iterative Spectral Algorithms; https://arxiv.

org/pdf/2602. 09530 [4] Automatic discovery of optimal meta-solvers for time-dependent nonlinear PDEs; https://arxiv. org/abs/2412.

00063 [5] Data-Driven Discovery of Interpretable Kalman Filter Variants through Large Language Models and Genetic Programming; https://arxiv. org/pdf/2508. 11703 [6] GRAFT-ATHENA; https://arxiv.

org/pdf/2605. 11117 Algorithm Discovery, Scientific Machine Learning (SciML), Artificial Intelligence, Optimization, Automated Algorithm Design, Differentiable Programming, Genetic Programming, Reinforcement Learning, Industrial Base For additional information and to submit your full proposal package, visit the DSIP Portal . Closes: Sept.

23, 2026 12:00 PM ET DoW STTR 2026 BAA | Release 5