The structure of technological learning: insights from water electrolysis for cost forecasting, policy, and strategy

1. Problem Statement

Forecasting the cost evolution of emerging clean technologies is critical for policy and investment but remains highly uncertain. While experience curves (or learning curves), which link cost declines to cumulative deployment, are the standard tool for such forecasts, their application to emerging technologies faces two major challenges:

1. Parametric Uncertainty: Learning rates (the percentage cost reduction per doubling of capacity) are difficult to estimate for nascent technologies.
2. Structural Uncertainty: There is no consensus on the structure of the learning model itself. Specifically, it is unclear how learning effects are distributed across:
   • Technology variants (e.g., Alkaline vs. PEM electrolyzers).
   • Geographic regions (e.g., Western vs. Chinese manufacturing ecosystems).
   • Cost components (e.g., mass-manufactured stacks vs. site-specific Balance of Plant/EPC costs).

The authors argue that standard models often assume a unified, global learning curve, which may obscure the realities of supply chain fragmentation, protectionism, and technological divergence, leading to flawed cost projections and policy strategies.

2. Methodology

The study uses water electrolysis as a case study due to its importance in decarbonization and the expectation of significant future cost reductions. The authors employ a component-based experience curve framework to project future costs under various structural assumptions.

Key Modeling Components:

• Cost Breakdown: Electrolysis project costs are divided into:
  • Stacks: The core technology (Alkaline [ALK] and Proton Exchange Membrane [PEM]).
  • Balance of Plant (BoP): Auxiliary systems (compressors, tanks, etc.).
  • Engineering, Procurement, and Construction (EPC): Site-specific design, permitting, and installation.
• Alternative Learning Structures: The authors compare three distinct structural assumptions for stacks and three for BoP/EPC:
  • Stacks:
    1. Shared Learning: All variants (ALK/PEM) and regions learn from a single global cumulative capacity.
    2. Technology-Fragmented: ALK and PEM learn independently (separate curves).
    3. Regionally Fragmented: Western and Chinese variants of each technology learn independently (four separate curves).
  • BoP/EPC:
    1. Fully Local: Learning occurs only within a specific region.
    2. Fully Global: Learning is driven by global cumulative capacity.
    3. Hybrid: BoP is treated as global (standardized design), while EPC is treated as local (site-specific execution).
• Scenario Analysis: The models are tested against a stylized deployment scenario (60–140 GW global capacity) to isolate the impact of structural assumptions from deployment volume differences.
• Tools: An interactive online dashboard was developed to allow users to explore these assumptions transparently.

3. Key Contributions

• Identification of Structural Uncertainty: The paper establishes that the structure of the learning model is a first-order determinant of cost forecasts, often outweighing the uncertainty in the learning rate parameter itself.
• Quantification of Fragmentation Effects: It provides a quantitative framework to assess how supply chain fragmentation (e.g., protectionism, "decoupling") and technological divergence alter cost trajectories.
• Differentiation of Cost Components: It highlights that different cost components (stacks vs. BoP vs. EPC) likely follow different learning dynamics (global vs. local), challenging the "one-size-fits-all" approach.
• Policy-Relevant Framework: The study moves beyond theoretical modeling to offer a decision-support framework for policymakers to stress-test strategies under multiple plausible futures.

4. Key Results

A. Electrolyzer Stack Costs

• Impact of Fragmentation: Under a shared learning model, PEM stacks remain more expensive than Alkaline stacks because they learn from a massive, aggregated historical base (including legacy ALK deployment). Under fragmented models (where PEM learns only from its own smaller, specific base), PEM costs drop much faster initially.
• Investment Requirements: To reach a target cost of $100/kW for Western PEM stacks:
  • Fragmented Model: Requires ~41 billion USD in cumulative investment and 287 GW of deployment.
  • Shared Model: Requires >890 billion USD in investment and 6,180 GW of deployment.
• Implication: Fragmented supply chains (regional isolation) drastically increase the capital and deployment required to achieve cost targets, making deep cost reductions harder to attain in a fragmented world.

B. BoP and EPC Costs

• Local vs. Global Dynamics:
  • EPC costs are highly sensitive to local conditions (labor, permitting). A local learning assumption implies that early movers in a region gain a permanent cost advantage, and other regions must "re-learn" the process.
  • BoP costs may benefit from global learning through standardization and integrated system designs.
• Regional Disparities: In the US, assuming fully local learning vs. fully global learning results in a projected cost difference of ~$215/kW for BoP/EPC by 2030. This translates to a $0.41/kg difference in the Levelized Cost of Hydrogen (LCOH).
• Combined Effect: When combining stack and BoP/EPC structural uncertainties, the total difference in projected US hydrogen costs can reach ~$0.75/kg, which is roughly 75% of the cost of natural gas-based hydrogen.

C. Learning Investment

The study introduces "learning investment" (cumulative capital required to reach a cost target). Results show that fragmented learning structures (regional or technological) significantly increase the "learning investment" required to reach the same cost targets compared to a unified global system.

5. Significance and Implications

• For Policy and Strategy:

  • Industrial Policy: Policies that restrict cross-border collaboration (e.g., export controls, local content mandates) may create "first-mover advantages" for domestic industries but will likely slow global cost declines by fragmenting learning.
  • Target Setting: Policymakers setting deployment targets based on a single global learning curve may underestimate the domestic deployment required to achieve cost competitiveness if learning is actually local.
  • Risk Management: Relying on a single model specification is dangerous. Strategies should be stress-tested against multiple structural assumptions (shared, fragmented, local, global).

• For Global Decarbonization:

  • If learning is local, the "trickle-down" cost reduction logic (where early adoption in wealthy nations lowers costs for developing nations) weakens. Developing nations may need to build local capabilities while technologies are still expensive, creating a higher barrier to entry.
  • Standardization: To accelerate global decarbonization, policy should encourage the standardization of plant designs and integrated systems to maximize the share of costs subject to global learning spillovers.

• Broader Applicability:

  • The framework applies to other emerging technologies with multiple variants and mixed cost structures, such as Small Modular Reactors (SMRs) and advanced nuclear, where the interplay between global manufacturing and local deployment is critical.

Conclusion

The paper concludes that structural uncertainty is not a secondary detail but a primary driver of forecast outcomes. For emerging technologies like electrolysis, assuming a unified global learning curve can lead to overly optimistic cost projections and flawed investment strategies. A robust approach requires evaluating multiple learning structures to understand the trade-offs between domestic capability building, supply chain resilience, and the speed of global cost reduction.