Designing IA for AI - Information Architecture Conference 2024
Analysis and Modelling of CMOS Gm-C Filters through Machine Learning
1. Analysis and Modelling of CMOS
Gm-C Filters through Machine
Learning
Malinka Ivanova
Technical University of Sofia
46th International Conference Applications of Mathematics in Engineering and
Economics, 7-13 June 2020
2. The aim
• A novel approach for design and analysis of
CMOS analog filters through utilization of
machine learning algorithms to be presented
3. Machine learning
• Machine learning is a part of artificial
intelligence and includes a set of algorithms
that allow systems to learn automatically
from data about them with aim to analyze,
optimize or predict their behavior
• This work is focused on supervised machine
learning where the learning is based on pairs
of input and output variables, training data set
and suitable machine learning algorithms
4. Machine learning algorithms
• Decision trees algorithms – for
decision making
– Identifying the correct class for every
training sample
– Explaining the decision
– Forming a set of rules
– Improving the tree model
5. Machine learning algorithms
• Decision trees algorithms
• Use divide-and-conquer technique
• Entropy H(X) –amount of information related to
an attribute
H(X)=1 – max value when P(x)=0.5
H(X)=0 - min value when P(X)=0 or 1
• Probability Pi=IXiI/IXI
– IXI – number of samples in the whole training set
– IXiI – number of samples in Xi
• Information gain – I(X, Y)=H(X)-H(X, Y)
6. Machine learning algorithms
• Decision trees algorithms
• Entropy calculation through percentages of samples in
class1 and samples in class2
H(X)=-Pclass1log2Pclass1-Pclass2log2Pclass2
• Several calculations for each attribute that divide the
training set X into subsets Xi:
– The entropy of each subset Xi
– The average entropy H(X, Y)=ΣPiH(Xi)
– Information gain I(X, Y)=H(X)-H(X, Y)
• The attribute with the maximum information gain is chosen
• The training set X is divided to the subsets Xi, characterized
with different attribute value
• If in a subset Xi all samples are from the same class, then a
leaf with the label of this class is created
7. Gm-C Filters
• Gm-C filters operates at higher frequencies
• Suitable for integration on chip (Bipolar, CMOS,
BICMOS technology)
• Electronically tunable
• Not so complex structure
• They are constructed on the operational
transconductance amplifiers (OTAs)
• OTA is a voltage controlled current source and
characterized with very big input and output
impedance and a given transconductance
8. Gm-C Filters
• The main researched structure is based on three-admittance
model
Y could be: g, sC, open
circuit, short circuit, g+sC
Deliyannis, Theodore L. et al "Single Operational
Transconductance Amplifier (OTA) Filters"
Continuous-Time Active Filter Design, 1999
+
-
gm
Ui
Y1
Y2
Y3
Uo
10. Machine learning for automated design
and analysis of electronic schemes
• Analysis – to receive the important scheme features based on its
topology
• Design – to obtain the scheme topology based on its main
parameters
• At this moment the topic about machine learning usage in support
of automation the design and analysis of electronic schemes is just
at the beginning
• Just few works are found:
• Peter A. Beerel and Massoud Pedram, Opportunities for Machine
Learning in Electronic Design Automation, 2018
• Elyse Rosenbaum, Machine Learning for Electronic Design
Automation, 2018
• Rod Metcalfe, Machine learning in EDA accelerates the design cycle,
2020
• Behzad Moradi and Abdolreza Mizaei, A New Automated Design
Method Based on Machine Learning for CMOS Analog Circuits, 2016
11. The proposed methodology
Analysis
Design
Existing
scheme
topology
1
Data
collection
about its
topology
and
functions
2
Data
model
prepara-
tion
3
Applying
machine
learning
algorithm
and data
training
4
Obtaining
the main
features
5
Scheme
parame-
ters
1
Data
collection
about the
parame-
ters
2
Data
model
prepara-
tion
3
Applying
machine
learning
algorithm
and data
training
4
Obtaining
the
scheme
structure
5
13. Analysis of Gm-C Filters
• Random Forest algorithm in RapidMiner Studio environment
Rules extraction:
IF the component2 is sC2 AND the component1 is g1 AND the component3 is g3+sC3 THEN
the filter type is band pass;
IF the component3 is g3 AND the component2 is g2 THEN the filter type is low pass.
@relation automatedanalysis
@attribute component1 {sc1, g1, inf, zero, g1+sc1}
@attribute component2 {sc2, g2, inf, zero, g2+sc2}
@attribute component3 {sc3, g3, inf, zero, g3+sc3}
@attribute filterorder {first, second}
@attribute filtertype {lp, hp, bp}
@data
sc1, inf, zero, first, lp
inf, g2, sc3, first, lp
sc1, inf, g3, first, lp
sc1, g2, g3, first, lp
g1, sc2, g3, first, hp
g1, g2, sc3, first, lp
sc1, g2, sc3, second, lp
sc1, g2, g3+sc3, second, lp
g1+sc1, g2, sc3, second, lp
g1, sc2, g3+sc3, second, bp
g1+sc1, sc2, g3, second, bp
14. Analysis of Gm-C Filters
• Random Forest algorithm
Rules extraction:
IF the component2 is g2 AND the component1 is inf THEN the filter is from first order;
IF the component3 is sC3 AND the component1 is g1+SC1 THEN the filter is from second
order
15. Dataset about Gm-C filters with CMOS
transconductor
Lee and
Cheng
2009 [20]
Zhang and
El-Masry,
2007 [21]
Qian et
al., 2005
[22]
Veeravalli
et al.,
2002 [23]
Solis-
Bustos et
al., 2000
[24]
Ramasamy
and
Venkataram
ani, 2018
[25]
Karami
and
Atarodi
, 2019
[26]
Supply
voltage
±1V ±1.8V ±1.5V ±1.35V ±1.5V 1.8V 1.5V
Filter order 5 3 5 2 6 2 1
Threshold
voltage
0.5V 0.43V 0.6V 0.9V 0.8V - -
Total
Harmonic
Distortion
48.6dB 45dB 49.5dB 45dB 50dB 40dB 44dB
Dynamic
range
50dB 45dB 57dB 70.5dB 60dB 50dB 68dB
Area 0.13 mm2 0.159
mm2
0.25 mm2 0.06 mm2 1mm2 0.3 mm2 -
16. Constructed decision trees
Rules extraction:
IF vdd is 1.5V AND thd is 49.5dB THEN the filter performance is verygood;
IF dr is 70.5dB THEN the filter performance is excellent;
17. Design of Gm-C Filters
• Random Forest algorithm in RapidMiner Studio environment
Rules extraction:
IF the component2 is g2 AND the transfer function is tf2 THEN the filter has to contain
two passive components
Rules extraction:
IF the gain is G10 THEN the filter has contain four passive components
18. Conclusions
• This work proposes a methodology for design and
analysis of Gm-C filters with a three-based machine
learning algorithm
• Several models are developed to verify the
methodology - datasets are prepared according to
published scientific results and mathematical
equations, data is trained in RapidMiner Studio
environment and Random Forest algorithm is applied
• The results point out that the classification machine
learning algorithm Random Forest precisely performs
the tasks regarding design and analysis of Gm-C filtes
Image is taken from: https://spectrum.ieee.org/image/MzEwMTk5OA.jpeg