Outlier Detection as Creative Method

Last October I ruined forty panels in a single afternoon.

Not by accident. I had been working with a batch of cold-rolled steel, preparing pieces for a commission, running my usual ferric chloride protocol. Controlled. Documented. The kind of process that yields predictable copper-brown surfaces with a satisfying patina, the surfaces people actually want to buy. Forty panels, identical treatment, identical steel. I could have done it in my sleep.

Instead I contaminated the bath. Salt water from the harbor, a handful of soil from the garden, vinegar from the kitchen, iron filings swept off the workshop floor. I poured it all in and watched the color of the solution go from amber to something murky and faintly green, like canal water in August. The smell was immediate: sulfur, brine, the mineral tang of wet earth. My hands were already stained brown from the ferric chloride, and now they picked up a grayish residue that wouldn’t wash off for days.

My wife came into the studio around four o’clock. She looked at the bath, looked at me, looked at the stack of steel blanks. “Those are for the Arnhem piece,” she said. Not a question.

I did know.

I submerged all forty panels. Left them overnight. The next morning the studio smelled like a tidal flat at low water, that heavy organic rot mixed with iron. I pulled the panels out one by one, rinsed them, laid them on the drying rack. Thirty-seven of them looked like exactly what you’d expect: ugly, uneven corrosion, muddy browns and dull grays, surfaces with no structure, no tension, nothing. Compost. Three weeks of preparation for the Arnhem commission, dissolved in canal water and garden dirt.

But three panels. Three out of forty. One had developed a crystalline blue-green formation along the lower edge, sharp geometric structures growing out of the chaos like frost on a window, except the color shifted to a deep violet where the crystals thinned. Another showed a surface I can only describe as geological: layered, striated, amber and black, as if someone had compressed ten thousand years of sediment into a few millimeters. The third was almost entirely black, but with a single streak of raw copper showing through, bright as a wound.

I stood in the studio holding that third panel, thirty-seven failures drying on the rack behind me, and thought: this is it. Not the panel. The ratio. Thirty-seven to three.

This is different from stumbling onto something beautiful by accident. I’ve written about that elsewhere, about the knowledge that lives in residuals, in what escapes your system. That’s passive. You compress reality into a method, and what falls through the cracks sometimes catches the light. But what I did with the forty panels wasn’t waiting for accidents. It was breeding them.

The question that won’t leave me alone: is this a method or a rationalization for not being able to control my materials?

A garden selects: you plant a hundred seeds knowing most will fail, and the ones that thrive in your particular soil, your particular light, become the garden. But a gardener has a vision of what should grow. I don’t. I’m planting seeds I’ve never seen, in soil I’ve deliberately poisoned, and then walking through the results trying to develop an eye for which mutations matter.

Kees, who runs the metal supply in Deventer, asked me once why I keep ordering steel in quantities that don’t match any project. “You’re either building a ship,” he said, “or you’re wasting money.” Both, I told him.

The algorithmic work follows the same logic. I run generative systems through thousands of iterations, parameter spaces so large that mapping them would take longer than my remaining years. I set the parameters at the edges of their ranges, where the algorithm starts to stutter. Feed it corrupted inputs. Let it argue with itself. What comes back is mostly noise, gray-green fields of pixel mush, the visual equivalent of static. But sometimes the system catches on a contradiction and won’t let go, and you get these structures, hard-edged, almost crystalline, colors that have no business being next to each other, vibrating on screen like something alive and in pain. Three percent. The ratio holds.

The word “detection” matters. I’m not generating the interesting failures. The contaminated bath generates them, the broken algorithm generates them. My job is detection. Walking through the failure garden with a magnifying glass, trying to distinguish the weed that’s merely a weed from the weed that’s a new species.

This requires a strange kind of attention. Not the focused attention of craft, where you know what you’re looking for and your hands know how to get there. Something peripheral. I’ve noticed that my best selections happen when I’m tired, slightly distracted, when the critical apparatus has gone slack. You have to stop trying to see. The outlier doesn’t survive scrutiny. It survives inattention.

Is there a difference between aesthetic discipline and neurosis? I ask myself this while sorting through algorithmic outputs at eleven at night, coffee going cold. Four images selected from maybe six hundred. Tomorrow I’ll look at the four again and probably keep one.

My friend Thomas, a data scientist in Amsterdam, tried to explain anomaly detection over dinner once. In his world, outliers are problems. Fraud, equipment failure, corrupted data. “You’d be a terrible data scientist,” he said. He’d been watching me push salt crystals around the tablecloth. “You’d keep all the fraud.”

I laughed. He wasn’t wrong.

What I do is the inverse of quality control. Quality control establishes a norm and eliminates deviation. I establish conditions for maximum deviation and then search the debris for significance. The art, if it is art, lives in the looking. In the capacity to stand in front of a hundred broken surfaces and feel, physically feel, which three carry something worth preserving. A tightening in the chest. A held breath.

Some days I contaminate the bath and pull out forty dead panels and none of them speak and I go home and eat dinner and feel like a fraud. Some weeks the algorithms produce nothing but sophisticated garbage and I wonder whether I’ve built an elaborate system for avoiding the harder work of actually making something with intention. The doubt is not an ornament on the method. It is the method.

So where does this leave me? Not quite an artist in any sense my father would have recognized. He painted. Oils on canvas, deliberate marks, the brush doing what the hand intended. I pour chemicals on metal and wait. A gardener of breakdowns, maybe. Or something that doesn’t have a name.

I keep coming back to that third panel from October. The almost-black surface with the copper streak. It hangs in the studio now, above the workbench. Every morning I look at it and I don’t know if it’s good.

The panel doesn’t care. It will continue oxidizing whether I watch it or not. The copper streak is already dulling, darkening, returning slowly to the field of black from which it came. In a year it will be gone. The outlier, re-entering the noise.

Perhaps that’s the only honest description of what I do: I build systems that fail, and then I pay very close attention.