Unilever's best mathematicians calculated the theoretically perfect nozzle and it still clogged — so the company stopped calculating and started breeding nozzles instead.

问题

a design problem is too complex for direct calculation to solve, even by genuine domain experts working from first principles

背景

换别人会怎么做

The natural response to a nozzle design problem was to bring in fluid-dynamics mathematicians to calculate the theoretically optimal geometry from physical first principles, the standard engineering approach of deriving the correct answer through expertise and equations rather than trial and error.

他们看到了什么

Unilever saw that the calculation-first approach had genuinely failed, not from insufficient expertise, but because the real fluid dynamics inside the nozzle were too complex to fully capture in a tractable calculation, meaning the problem itself exceeded what direct analytical derivation could solve regardless of how skilled the mathematicians were. Rather than continuing to refine the theoretical model, the fix was abandoning calculation entirely and treating nozzle design as an evolutionary process, generating variants, testing them against the real problem, and repeatedly selecting and mutating whichever performed best, reaching a working solution without ever needing to fully understand it analytically.

那一手

The biologists treated nozzle design as an evolutionary process rather than a calculation: starting from an existing nozzle, they generated ten random variations, tested all ten against the real clogging and granule-consistency problem, kept whichever variant performed best, generated ten new random variations of that winner, and repeated the cycle for 45 generations.

为什么管用

Generating ten random variations of an existing nozzle, testing all ten against the actual clogging and granule-consistency problem, and repeating that cycle for 45 generations let the process discover what worked through direct empirical feedback rather than through a model of the underlying physics that had already proven too simplified to capture reality. Because each generation's selection was based on real, measured performance against the actual problem rather than a theoretical prediction, the process could converge on solutions no calculation would have derived, including a shape strange enough that no one on the team could fully explain in physical terms why it outperformed the mathematically 'optimal' design. This is why after 45 generations and 449 failed variants, the process produced a nozzle over 100 times more efficient than the original, a result achieved specifically because the method didn't require understanding the mechanism, only measuring the outcome and iterating toward better performance.

值了多少

After 45 generations and 449 failed variants, the process converged on a nozzle design over 100 times more efficient than the original — reliably avoiding clogging and producing consistent granules — even though the resulting shape was strange enough that no one on the team could fully explain in physical terms why it worked better than the mathematically 'optimal' design the fluid dynamicists had calculated.

什么时候会失灵

The mechanism depends on having a fast, cheap, and reliable way to test each generated variant against the real problem, evolutionary search requires running many iterations, and a problem where each test is prohibitively slow, expensive, or risky to run in the real world would make 45 generations of trial and error impractical regardless of how well the method works in principle. It also depends on the problem space actually being navigable through incremental variation and selection, a landscape with no smooth path from a starting point to a good solution, where small variations produce no meaningful signal about direction, would leave the evolutionary process wandering without making the kind of steady progress that converged on Unilever's nozzle. And producing a working solution without understanding why it works carries a real practical cost: without a physical explanation for the winning design, engineers have less ability to predict how it will perform under conditions not covered by the original tests, or to adapt it confidently to a related but different problem, meaning the approach trades interpretability for results in a way that limits how far the solution's lessons can be generalized elsewhere.

后来呢

The Unilever nozzle case is a standard teaching example in innovation and complexity science for the limits of theory-first design on genuinely complex problems, and for evolutionary or 'generate-test-select' optimization as a viable alternative when direct calculation fails — the same underlying method now underlies genetic algorithms in engineering design, machine-learning architecture search, and pharmaceutical compound optimization.

资料来源

1. [1]Sex and Soap Powder, Trial and ErrorThe Squawk Point, 2016squawkpoint.com
2. [2]What 449 'Failures' Can Teach Us About SuccessSam Thomas Davies, 2020samuelthomasdavies.com

收下它

同一路聪明

• 1979 · NASA / United Airlines (Crew Resource Management)Aviation stopped losing planes to captains who wouldn't listen by making it a subordinate's job to make them
• 1958 · Bank of America (the 'Fresno Drop')No bank knew if ordinary people would actually use unsecured credit — so instead of running a pilot with volunteers, Bank of America just mailed 60,000 live credit cards to an entire city and watched what happened.
• 1948 · A&P (with USDA)Convicted of monopoly abuse, A&P rebuilt its reputation by giving away a better chicken instead of buying better ads