Computing within Materials: Self-Adaptive Materials and Self-organizing Agents

Overview

Material Informatics: Material-integrated Computing

Smart Materials

Smart Materials: Fusion of Sensorial- and Adaptive Materials with Information Processing

Sensorial Material

A material or structure with integrated sensing and data processing (ICT)

Adaptive Material

A material or structure with integrated sensing, data processing and actuation that can control and change material or structure properties

McEvoy, 2015 [1]

Smart Materials

In our understanding, a smart material provides the following major features:

1. Perception using various kinds of sensors, e.g., measuring of strain, displacement, temperature, pressure, forces;

2. Changing of local material and structure properties by actuators, e.g., stiffness or damping variation;

3. Integrated Information and Communication Technologies (ICT);

4. Distributed Approach: Local sensor processing and actuator control - Global cooperation and coordination.

5. Robustness and Self-Organization

Smart Materials

Material Informatics: Computing within Materials

• Traditionally computation is separated from sensing and control

• Smart Materials poses the tight coupling of computation, communication, sensing, and control with loosely coupled nano computers

• Algorithmic scaling and distribution are required

  

Computing Power and Efficiency

• A normalized computing efficiency of a computer (considering only the data processing unit) can be defined by

(1)

A

Chip Area in square millimetres

C

Computing Power in (Integer) Mega Instructions Per Second (MIPS)

P

Electrical Power in Watt

• The computing efficiency can be used to compare different computers and devices, i.e., giving a scaling factor:
  s=ε_x/ε_y

Technologies

Existing “Nano” Computers

• Smart Dust → Millions of loosely coupled nano computers, e.g.,
  • embedded in materials
  • scattered on surfaces
  • dispersed in liquids, foils, ..
  • about 10mm³ volume

Comparison of Computers

Optimization and Adaptation

Physical Model

• A component composed of a Smart Material consists of:

  • Mass (body) nodes or material regions
  • Interconnection elements with parameterizable properties

Multi-body Physics Model

Solid body mass nodes with springs connecting nodes:

• A spring is an actuator with two optimization target variables:
  Stiffness s, Damping d

• Each spring is a strain sensor delivering the sensor value σ for the computation of the observation variable

Optimization Goals

Reduction of global and local stress, strain, or forces of arbitrariely shaped components under varying load situations

Control and Optimization Cycle

1. Perception using sensors

2. Comparison of local and global observation variables

3. Modification of material parameters by actuators

Algorithms

do with x ∈ {ε,σ,U}
 X:=0; ∀n∈ℕ do X := X + xn
 X := X/|ℕ|
 ∀n∈ℕ do
  rn := kx(xn/X, sn) 
  sn := sn * rn
until |Err| < Err0

do with x ∈ {ε,σ,U}
 ∀Si∈𝕊 do
  Xs,i:=0;
  ∀n∈Si do Xs,i := Xs,i + xn
  Xs,i := Xs,i / |Si|
  ∀n∈Si do
    rn := kx(xn/Xs,i, sn)
    sn := sn * rn
until |Err| < Err0

do with x ∈ {ε,σ,U}
 ∀{ni,nj ∈ N | i≠j ∧
  |pos(ni)-pos(nj)|=1} do
  if xi/xj < 1 ∧
     si-Δs > s0 ∧
     sj+Δs < s1 then
    si := si - Δs
    sj := sj + Δs
  end if
always

Multi-Agent Systems

Sensor Data Distribution

• Each node agent sends its sensor values to neighbour nodes within a range of radius 1

• This completes the set of sensor each node requires

• Remote signals are used for sensor distribution using &Delta-distance routing

Distributed Observable Computation

• Global observable computation in a large-scale distributed network is expensive and difficult (failures)

  • Random walk and directed diffusion can be used to approximate a global observable

• Segmented approach requires network segmentation

  • Without central instance difficult;

  • Instead a floating segment window is placed around each node

  • Each node has observable from all direct neighbours (North, South, West, East, Up, Down)

  • A chained distribution of data is used in each segment (N nodes ↔ N segments!)

  • Observable values with distance r=1 and r=2 are collected by each node to compute region observable

Distributed Observable Computation

Distance r=1

Observable from direct neighbours

Distance r=2

Direct neighbours deliver also values form their neighbours (opposite to request direction)

Distributed Adaptation

• The global and the segment algorithm need no further distribute coordination
• After the region observable (global or segment) is computed the actuators (springs) of each node can be modified basing on ratio of the local and the region observable value
• Neighbour negotiation approach do not require a region observable

Simulation

Multi-domain Simulation Framework

Simulation Example

• Device under Test: Plate (8x5x3 nodes), large hole, external load

Simulation Results

• Global and segment optimization achieves 40% decrease of total strain and maximum strain energy of mass elements using a linear correction function

• Neighbour negotiation is simple but not as efficient as segment approach

Conclusions and Outlook

• Smart Materials poses the tight coupling of computation, communication, sensing, and control with loosely coupled nano computers

• Algorithmic scaling and distribution are required

• Distributed information processing paradigma: Multi-agent Systems

• Multi-domain simulation enables the development and evaluation of different optimization strategies for smart adaptive materials and structures

• The SEJAM simulator enables the simulation and analysis of coupled physical and computational systems

• Global and segment optimization achieves 40% decrease of total strain and maximum strain energy of mass elements using a linear correction function

References

[1] M. A. McEvoy and N. Correll, “Thermoplastic variable stiffness composites with embedded, networked sensing, actuation, and control,” Journal of Composite Materials, vol. 49, no. 15, 2015.