著者

米倉 一男 服部 均 斉藤 弘樹 鈴木 克幸

出版者

一般社団法人 人工知能学会

雑誌

人工知能学会全国大会論文集 第34回 (2020)

巻号頁・発行日

pp.3H5GS305, 2020 (Released:2020-06-19)

---

機械設計では以前から機械学習を用いた最適設計手法が用いられてきた。近年の深層学習の進歩に伴って、これまでよりも推定精度が向上し、より幅広い設計プロセスに応用できるようになった。 本講演では深層学習を機械設計プロセスに応用するアプローチを三つ紹介する。すなわち回帰モデルによる性能推定、生成モデルによる形状生成、強化学習による形状修正である。これらの方法を、深層学習や深層強化学習wお用いた場合とそれ以外の方法を用いた場合を比較すると、深層学習等を用いた場合のほうが高精度で推定ができることを示す。これらのアプローチはデータを基にしているという意味でたデータ駆動型設計と呼べる。

著者

西津 卓史 谷次 智弥 竹澤 晃弘 米倉 一男 渡邊 修 北村 充

出版者

一般社団法人 日本機械学会

雑誌

日本機械学会論文集 (ISSN:21879761)

巻号頁・発行日

vol.83, no.855, pp.16-00581-16-00581, 2017 (Released:2017-11-25)

参考文献数

19

被引用文献数

2

---

Structure can get various mechanical characteristics by applying periodic structures as typified by lattice structures. Lattice structures are generally used inside the structural member in order to reduce the weight. One advantage of lattice structures is that we do not need to change the whole structural shape when we replace the solid part of a component with the lattice structures. Another advantage is the lightness of the weight, and hence it is important to design a high performance lattice shape with low weight. However, a framework for development of micro lattice structures considering both stiffness and weight has not been established. Thus, we propose a method for designing and producing micro lattice structures. We use a topology optimization method for a designing methodology. Topology optimization is an effective method in designing high performance lattice structure since topology optimization allows us to change the topology and to design a complicated shape. We use a metal additive manufacturing (AM) machine for producing the optimal lattice structures. AM allows us to produce a complicated structure which removal and forming manufacturing cannot produce. We use a bulk modulus as the objective function since it is one of the important mechanical characteristics in design. In this research, we use a homogenization method to compute the bulk modulus. Objective function was modified so that isotropy of the optimal shape is retained when the solution is updated. In addition, structures produced by AM need holes so that internal metal powder can be removed. Hence, we defined the design domain so that the optimal structure becomes open cell structure. Then, high bulk modulus shapes were derived using topology optimization. The lattice structures were produced by metal AM machine after being modified for production.

著者

米倉 一男 寒野 善博

出版者

一般社団法人 日本機械学会

雑誌

日本機械学会論文集 (ISSN:21879761)

巻号頁・発行日

pp.15-00337, (Released:2015-12-04)

参考文献数

24

---

We propose a Newton-gradient-hybrid optimization method for fluid topology optimization. The method accelerates convergence and reduces computation time. In addition, the fluid-solid boundaries are clearly distinguished. In the method, the optimization process and flow computation are executed concurrently. The flow computation utilizes the lattice Boltzmann method (LBM), and the optimization algorithm partly utilizes a Hessian matrix. Due to the formulation of LBM and the optimization algorithm, the Hessian matrix is a diagonal matrix. Since the optimization problem is nonconvex problem, the Hessian matrix is not generally positive semidefinite. Hence, we employ a gradient method for a component whose corresponding Hessian matrix elements are negative. We compare the optimization results with those of conventional gradient method and show that the convergence is accelerated and the fluid-solid boundaries are clearly distinguished.