採用機器學習方法並結合聚類算法，從而獲得淬火局域的無序分佈。
淬火無序現象，是人們對各種材料（如FCC和BCC晶體、無定形固體）和地震地質斷層突發塑性事件或材料爆裂噪聲事件進行觀察而得到認識的。爆裂噪聲可由隨機-場模型或界面定位模型加以解釋，涉及均勻固體的彈性、局域淬火無序，以及微觀狀態空間允許的不均勻和隨機分佈。然而局域淬火無序卻一直難以測量。美國西弗吉尼亞大學的Stefanos Papanikolaou教授採用無監督機器學習方法並結合聚類算法，以期從具有爆裂噪聲隨時間演化行為的應力-應變曲線中獲得淬火、局域的無序分佈。該方法在兩種爆裂噪聲模型中能成功實現數據的聚類和分類，並從鎳微柱單軸壓縮實驗的數據中成功得到了淬火無序的分佈。這是典型的時間局域可觀察參量（如突發事件大小/持續時間）途徑所無法企及的。作者將這一方法記作時間序列-機器學習法。若將這些淬火無序分佈的識別及分類擴展到不同材料、加載模式和樣品加載歷史中，將有助於建立隨機屈服分佈的數據庫，進而改進多尺度力學模型。該文近期發表於npj Computational Materials4:27(2018); doi:10.1038/s41524-018-0083-x。英文標題與摘要如下，點擊“”可以自由獲取論文PDF。
Learning local, quenched disorder in plasticity and other crackling noise phenomena
Stefanos Papanikolaou
When far from equilibrium, many-body systems display behavior that strongly depends on the initial conditions. A characteristic such example is the phenomenon of plasticity of crystalline and amorphous materials that strongly depends on the material history. In plasticity modeling, the history is captured by a quenched, local and disordered flow stress distribution. While it is this disorder that causes avalanches that are commonly observed during nanoscale plastic deformation, the functional form and scaling properties have remained elusive. In this paper, a generic formalism is developed for deriving local disorder distributions from field-response (e.g., stress/strain) timeseries in models of crackling noise. We demonstrate the efficiency of the method in the hysteretic random-field Ising model and also, models of elastic interface depinning that have been used to model crystalline and amorphous plasticity. We show that the capacity to resolve the quenched disorder distribution improves with the temporal resolution and number of samples.
閱讀更多 知社學術圈 的文章
