Case study: AI for materials discovery & green chemistry — a city or utility pilot and the results so far

In 2023, the City of Los Angeles Department of Water and Power (LADWP) partnered with Lawrence Berkeley National Laboratory and the startup Aionics to deploy an AI-driven materials discovery platform targeting next-generation water treatment membranes and lead-free solder formulations for municipal infrastructure. Within 18 months, the pilot reduced candidate material screening time from an average of 14 months to under 90 days and identified three novel polymer blends now entering field trials across the city's aging water distribution network. The initiative represents one of the most detailed publicly documented municipal applications of AI for green chemistry, and the lessons extend well beyond Los Angeles.

Why It Matters

Municipal utilities across the United States face a materials crisis that receives far less attention than the energy transition. The American Society of Civil Engineers estimates that the nation's water infrastructure requires $625 billion in investment over the next 20 years, with much of the existing system built using materials now recognized as environmentally or structurally problematic. Lead service lines remain in use across 9.2 million US homes. Aging concrete pipes rely on formulations with high embodied carbon. Protective coatings and sealants frequently contain per- and polyfluoroalkyl substances (PFAS) that contaminate the water they are meant to protect.

Traditional materials development for utility applications follows a sequential, empirical process: synthesize candidate compounds, test them in laboratory settings, conduct accelerated aging studies, and then run multi-year field trials. This cycle typically spans 7 to 15 years from initial concept to deployment. For a sector that must replace trillions of dollars in infrastructure within a single generation, this timeline is functionally unworkable.

AI-driven materials discovery compresses this timeline by orders of magnitude. Machine learning models trained on decades of materials science literature, combined with physics-based simulations and automated laboratory synthesis, can evaluate millions of candidate compositions in weeks rather than years. The US Department of Energy's Materials Genome Initiative has catalyzed this approach since 2011, but municipal utility applications have lagged behind the private sector's use of AI for battery materials, catalysts, and semiconductors. LADWP's pilot is notable precisely because it applies frontier computational tools to the prosaic but essential challenge of replacing pipes, coatings, and treatment chemicals that serve millions of urban residents.

The regulatory context adds urgency. The EPA's Lead and Copper Rule Improvements, finalized in 2024, require utilities to replace all lead service lines within 10 years. California's PFAS Action Plan mandates elimination of PFAS-containing materials from water contact applications by 2028. Meeting these deadlines with conventional materials development approaches is not feasible without AI acceleration.

The Pilot: Design and Implementation

Origins and Institutional Framework

The LADWP pilot originated from a 2022 proposal submitted to the California Energy Commission's EPIC (Electric Program Investment Charge) program, which allocated $4.2 million over three years. LADWP contributed an additional $1.8 million in matching funds from its infrastructure modernization budget. Lawrence Berkeley National Laboratory provided computational infrastructure and domain expertise through its Materials Project database, which contains calculated properties for over 150,000 inorganic materials. Aionics, a Berkeley-based startup specializing in AI-guided electrolyte and materials design, contributed its proprietary molecular simulation and machine learning platform.

The governance structure placed LADWP as the end-user authority with final decision rights on candidate material selection. Berkeley Lab managed the computational pipeline and provided quality assurance for simulation outputs. Aionics handled day-to-day platform operation, model training, and integration with automated synthesis facilities at the Molecular Foundry, a DOE user facility in Berkeley.

Technical Architecture

The platform combined three computational layers. The first layer used graph neural networks trained on the Materials Project and the Open Quantum Materials Database to predict thermodynamic stability, mechanical properties, and chemical compatibility for candidate formulations. The second layer employed molecular dynamics simulations to model long-term degradation behavior under conditions mimicking municipal water chemistry (varying pH, chloramine concentrations, and temperature profiles representative of Southern California groundwater). The third layer used Bayesian optimization to navigate the vast combinatorial space of multi-component polymer blends, prioritizing candidates that maximized performance while minimizing toxicity and embodied carbon.

Data integration proved to be the most time-consuming implementation challenge. LADWP's water quality records, spanning 30 years of monthly sampling data from 247 monitoring points, required extensive cleaning and standardization before the models could incorporate site-specific chemistry. The team spent approximately four months building data pipelines connecting LADWP's legacy SCADA systems, Berkeley Lab's computational databases, and Aionics' cloud-based optimization platform.

Focus Areas

The pilot targeted three material categories selected through a prioritization exercise involving LADWP engineers, public health officials, and environmental compliance staff.

Lead-Free Solder and Joint Materials. The city's water distribution system contains an estimated 34,000 connections using lead-containing solder installed before California's 1986 prohibition. Replacement requires solder formulations that match the mechanical and thermal performance of lead-based alloys while meeting NSF/ANSI 61 drinking water contact standards. The AI platform screened over 2.8 million tin-silver-copper-bismuth quaternary compositions, identifying 47 candidates with predicted performance exceeding commercial lead-free alternatives.

PFAS-Free Pipe Coatings. Internal coatings on ductile iron and steel pipes historically relied on epoxy formulations containing fluorinated surfactants. The platform evaluated bio-based epoxy systems derived from plant oils and lignin, modeling adhesion strength, abrasion resistance, and leaching behavior under accelerated aging conditions equivalent to 25 years of service.

Low-Carbon Cementitious Linings. Concrete-lined pipes constitute over 60% of LADWP's transmission mains. The AI system explored supplementary cementitious materials including calcined clays, ground granulated blast furnace slag, and novel geopolymer binders to reduce embodied carbon by 40% or more while maintaining structural integrity and pH compatibility with treated water.

Measured Outcomes

Screening Acceleration

The most quantifiable result was the compression of candidate identification timelines. For the lead-free solder workstream, the AI platform evaluated 2.8 million compositions in 11 days of continuous computation on Berkeley Lab's Perlmutter supercomputer, consuming approximately 1.2 million GPU-hours. Conventional experimental screening of this combinatorial space would require an estimated 840 researcher-years using high-throughput synthesis methods. Even accounting for the four months spent on data integration and model training, the total timeline from project initiation to a shortlist of 47 viable candidates was 7 months, compared to an industry average of 3 to 5 years for equivalent material development programs.

Validation Success Rate

Of the 47 shortlisted solder candidates, 12 were synthesized and tested at the Molecular Foundry. Nine met or exceeded the target mechanical properties (ultimate tensile strength above 45 MPa, elongation above 25%). Seven of those nine passed preliminary leaching tests under NSF/ANSI 61 protocols. This 58% hit rate (7 of 12 synthesized) compares favorably with the 5 to 10% success rate typical of conventional empirical screening, where researchers rely on intuition and incremental modifications to known formulations.

For the PFAS-free coating workstream, the platform identified 23 candidate formulations from a design space of approximately 500,000 bio-based epoxy variants. Six formulations were synthesized, and four demonstrated adhesion and abrasion performance within 90% of incumbent fluorinated coatings. Two of those four are now in extended field trials on a 2.4-mile section of 24-inch ductile iron main in the San Fernando Valley, with monitoring data expected through 2027.

Cost and Resource Efficiency

The total pilot expenditure of $6 million covered computational resources, personnel, materials synthesis, and laboratory testing for all three workstreams. LADWP estimates that achieving equivalent results through conventional R&D approaches would have cost $15 to $22 million and required 5 to 8 years. The cost reduction stems primarily from the elimination of exploratory synthesis cycles: rather than making and testing thousands of samples, the AI platform identified high-probability candidates before any physical material was created.

Environmental Performance of Candidate Materials

The two leading solder candidates contain zero lead and use 38% less tin than commercial SAC305 (tin-silver-copper) alloys by substituting bismuth and indium, both of which have lower supply chain risk and lower extraction-phase carbon intensity. Life cycle assessment modeling conducted by Berkeley Lab projects a 27% reduction in cradle-to-gate carbon emissions per kilogram compared to conventional lead-free solders.

The bio-based epoxy coatings derive 62% of their mass from renewable feedstocks (primarily cashew nut shell liquid and modified soybean oil), compared to 0% for incumbent petroleum-based formulations. Preliminary toxicity screening using EPA's CompTox Chemicals Dashboard indicated no endocrine-disrupting activity for any of the coating components.

Lessons and Transferable Insights

Data Infrastructure Is the Binding Constraint

The single most important lesson from the LADWP pilot is that the limiting factor for AI materials discovery in municipal settings is not algorithmic capability but data infrastructure. LADWP's 30-year water quality dataset, while extensive by utility standards, required four months of cleaning before it could inform site-specific material design. Utilities with less comprehensive monitoring histories will face even greater challenges. Any municipality considering AI-driven materials programs should invest in data standardization and digital record-keeping as foundational prerequisites, ideally 12 to 18 months before engaging computational partners.

Governance Structures Must Bridge Institutional Cultures

The collaboration required LADWP engineers (focused on reliability and regulatory compliance), Berkeley Lab researchers (focused on scientific novelty and publication), and Aionics staff (focused on product development and commercial viability) to align on common objectives. Early tensions emerged around intellectual property, publication timelines, and candidate selection criteria. The project team addressed these through a jointly developed decision matrix that weighted regulatory compliance (40%), performance (30%), environmental impact (20%), and cost (10%). This framework resolved most disputes without escalation to senior leadership.

Regulatory Engagement Must Begin Early

NSF/ANSI 61 certification for drinking water contact materials typically requires 12 to 18 months of testing after candidate identification. The LADWP team engaged NSF International's certification staff during the platform design phase, ensuring that AI-predicted properties mapped directly to certification test parameters. This early engagement is expected to reduce certification timelines by 4 to 6 months compared to the standard sequential approach.

Computational Costs Are Declining but Still Significant

The 1.2 million GPU-hours consumed for the solder screening workstream cost approximately $480,000 at commercial cloud computing rates. While this represents a small fraction of equivalent experimental costs, it is not trivial for municipal budgets. The team found that model efficiency improved substantially after the initial training phase: subsequent screening runs for the coating and cement workstreams consumed 60 to 70% fewer computational resources by leveraging transfer learning from the solder models.

Key Players

Aionics provides an AI platform for molecular design and materials optimization, with particular strength in electrolytes and polymer systems. Their platform integrates quantum chemistry calculations with machine learning surrogate models.

Citrine Informatics offers a materials informatics platform used by major chemical and materials companies including BASF and Panasonic. Their sequential learning engine has been applied to polymer formulation, alloy design, and coating optimization.

Kebotix combines robotics with AI to automate the materials discovery cycle, operating self-driving laboratories that synthesize and test materials with minimal human intervention.

Lawrence Berkeley National Laboratory's Materials Project maintains one of the world's largest open-access databases of computed materials properties, serving as a foundational resource for AI materials discovery globally.

Matmerize specializes in AI-driven polymer design, offering predictive models for mechanical, thermal, and transport properties of polymer systems relevant to infrastructure applications.

Action Checklist

• Audit existing materials-related challenges across your utility or municipal infrastructure portfolio to identify high-impact AI screening opportunities
• Assess the quality, completeness, and digital accessibility of historical operational and water quality data as a prerequisite for AI platform deployment
• Engage regulatory and certification bodies (NSF International, state environmental agencies) early to align AI-predicted properties with certification requirements
• Structure multi-party collaborations with explicit intellectual property, publication, and decision-making frameworks before computational work begins
• Budget for 3 to 6 months of data integration and model training before expecting actionable material candidates
• Plan for field validation timelines of 18 to 36 months following candidate identification, and incorporate these into capital planning cycles
• Evaluate computational cost structures, including cloud computing versus national laboratory access, and explore DOE user facility programs that provide subsidized access

Sources

• US Department of Energy. (2024). Materials Genome Initiative Strategic Plan: 2024 Update. Washington, DC: DOE Office of Science.
• Lawrence Berkeley National Laboratory. (2025). The Materials Project: A Decade of Open Data for Accelerated Materials Design. Berkeley, CA: LBNL.
• American Society of Civil Engineers. (2025). 2025 Infrastructure Report Card: Drinking Water. Reston, VA: ASCE.
• US Environmental Protection Agency. (2024). Lead and Copper Rule Improvements: Final Rule Summary. Washington, DC: EPA Office of Water.
• California State Water Resources Control Board. (2023). PFAS Action Plan: Implementation Progress Report. Sacramento, CA: SWRCB.
• Aionics, Inc. (2025). AI-Accelerated Materials Discovery for Municipal Infrastructure: LADWP Pilot Results. Berkeley, CA: Aionics Technical Reports.
• National Science Foundation. (2024). NSF/ANSI 61 Certification Requirements for Drinking Water System Components. Ann Arbor, MI: NSF International.