通过机器学习，我们将跳过100年

Every part-making process is subject to variables that affect its outcome, but with additive manufacturing (AM), there is a difference of degree. In machining, for example, we know the impact of choices of tooling, cutting fluid, feed rate and so on, and in molding we know the effect of characteristics of the material and mold design. But when it comes to an additive process such as selective laser melting in metal, there are more variables and more combinations of variables affecting the form and properties of the part than we have yet recognized, let alone mastered.

但是，添加剂将受益于另一个重要区别：时机的差异。

Other manufacturing processes came of age well in the past. Additive is maturing during a time of advanced computing power. As a result, it appears likely that machine learning—the theme of this month’s issue ofAdditive Manufacturing- 将被证明是进步方式的主要因素。

Specifically, machine learning will enable the speed of AM’s advance. The term does not imply a type of calculation humans can’t do, but instead a volume of calculations. A billion different relationships among many inputs and outputs can be explored rapidly through computation, whereas just a fraction of the same exploration might take human beings working on their own something like a century to carry out.

我们知道这一点，因为过去，我们别无选择，只能花那个世纪。例如，过去一年的新型加工设计可能已经开始研究前十年或更长时间。我们的知识以这种方式，甚至在世代上进行了缓慢的发展，研究人员通过实验及其对所见事物的推断，专门探索过程变量。AM的早期成功也是如此。

在本月的报道中，可以瞥见知识和成功的新方式。本月的三篇特征文章（与过程控制,材料和AI驱动的决策) describe how machine learning is being used as a tool to more rapidly advance our mastery of AM.

And a tool is just what machine learning is, says Bryce Meredig, co-founder of Citrine Informatics, the firm applying this tool as part of the research discussed在本文中. Machine learning is a different kind of tool than anything manufacturers are accustomed to, but it is one they will not only come to understand, but also make better through their efforts. He uses the term “artificial intelligence” (AI), and to him, “artificial” captures both the tool’s power and its limitations.

“AI is really good at taking a billion possibilities and short-listing them to ten,” he says. “But humans have important pieces of understanding that are too subtle to incorporate into computation.” Thus, only a human can look at a set of possibilities produced by machine learning and see something like,Options one and two are not feasible, but option three is a promising way to go.

Success will lead to success, he says. Progress will accelerate because mathematical relationships will be found that point to physically meaningful relationships. Rules will be found. And when a rule is discovered—that is, do X in selective laser melting of Inconel 718 and it reliably leads to Y—then that knowledge can be incorporated into the model, allowing subsequent machine learning to accept the rule as known and direct its power elsewhere.

Humans will find the rules, he notes. That much will remain as true as it ever was. As computation identifies promising connections between inputs and outputs, human experts will judge which findings are noise and which make sense. It is simply the speed of discovery that will change, but this change will be profound, and we will even see it accelerate. That is, as people continue to perform the crucial last step of recognizing the value in AI’s findings, he says, that very contribution will also allow the pace of those findings to increase.