* RF Power Amp Driver

 RF Power Amp Driver  (one stage) 

PA Lessons, PA Verified by computing the exact KCL residual at its converged solution, not because of blow up of MNA

SMITH Chart,  Explain Smith Chart

Design Work Flow

Design matching network

artifact slides

corrected; Angelov model  (share)

PA revision with physical accuracy

what has been changed (shifting not chasing, good)

MNA

Optimizer Apply MNA

what has been changed surface, reality

artifact

artifact 

Optimizer Apply Harmonic (close the loop)

change v9 MNA vs V21 Harmonic

Doherty architecture

Gemini 3.0 Pro

Opus 4.7 Adaptive

Doherty

Doherty PA 的真實規格範圍

PAPR、Back-off、與 Doherty 的存在理由(share)

v22

change documentation

visualization (pt A&B by tuning v22 2.2*Vdd, 2*Vdd too tight)

learning roadmap

v23 self-heating loop

v24 MNA for S parameters

v25* IMN+OMN, brute force+ NM Opti.

 corrected, v26* IMN+OMN, brute force+ CMA-ES (v23-v26 change log)

block diagram artifact

RF PA driver with 10 dBm output (not PA itself) (code, artifact, realistic spec, not just renaming)

Vdd @N3E nominally 0.75, max 0.8 V. The simulator default 3.3 V, too high

fix 2. thick ox 1.8v N4P corrected, cascode (option single also works.) HB, Newton compute -- educational, no novelty

Dev Log

Note: The Angelov model is the wrong device model for CMOS FinFETs (use BSIM-CMG or at least a FinFET-fitted surface-potential model), "N3E-RF" isn't a real process offering, and the Apple C1 framing implies an integrated Doherty PA that doesn't exist in that chip.

EX#7 Mathematical Optimization

 Assignment Structure 作業架構

md version 

Deadline:  Next Saturday at 23:59 (one more week) 期中考緣故，順延一周

Send all the share links to  me chang212@gmail.com by email with subject EX#7 [your id, your name]

Students choose from three progressively harder assignments. Each builds on the interactive "Optimization in Hyperspace" 3D visualizer, which demonstrates five algorithms navigating a fitness landscape.

同學從三份難度遞增的作業中逐一完成。每份皆以互動式「超空間最佳化」3D 視覺化工具為基礎，展示五種演算法在適應度地形上的行為。

Problem 1 — Conceptual Understanding 概念理解

• Task 任務： Written reflection (300–500 words) on how each algorithm behaves on the landscape — local vs. global optima, exploration vs. exploitation.
• 撰寫心得（300–500字），描述各演算法在地形上的行為差異——區域最佳解 vs. 全域最佳解、探索 vs. 開發的取捨。
• Deliverable 繳交物： md (1–2 pages)
• Difficulty 難度： ★☆☆

Problem 2 — Mathematical Analysis 數學分析

• Task 任務： Implement gradient descent from scratch on a multimodal 2D function f(x,y) = sin(x)·cos(y) + 0.1(x² + y²). Test 3 starting points × 3 learning rates. Plot the surface + convergence paths.
• 自行實作梯度下降法，應用於多峰 2D 函數。測試 3 個起點 × 3 種學習率 (α = 0.01, 0.1, 0.5)，繪製曲面與收斂路徑。
• Deliverable 繳交物： Code + 1-page report in artifact
• Difficulty 難度： ★★☆

Problem 3 — Comparative Algorithm Design 演算法比較實驗

• Task 任務： Compare Gradient Descent vs. Simulated Annealing (exponential cooling T(t) = T₀·γᵗ). Run 50 trials each from random starts; vary T₀ and γ; produce success-rate table + plot.
• 比較梯度下降 vs. 模擬退火法（指數冷卻排程）。各執行 50 次隨機起點實驗，調整 T₀ 與 γ，製作成功率表格與圖表。
• Deliverable 繳交物： Code + 1–2 page analysis with table & plot in artifact
• Difficulty 難度： ★★★

---

Five Algorithms in the Visualizer 視覺化工具中的五種演算法

Algorithm 演算法	Type 類型	Key Trait 特性
Gradient Descent 梯度下降	Local 區域	Follows steepest slope; gets trapped in local optima 沿最陡方向走，易陷入區域最佳解
Nelder-Mead 單純形法	Local 區域	Derivative-free simplex; can stall near local optima 無需導數，但仍可能停在區域解
A* Search A*搜尋	Heuristic 啟發式	Graph-based, uses heuristic to guide search 圖形搜尋，以啟發函數引導
Simulated Annealing 模擬退火	Global 全域	Accepts worse moves probabilistically; escapes traps 以機率接受較差解，可跳脫陷阱
Global Optimization 全域最佳化	Global 全域	Systematic global search 系統性全域搜尋

---

Grading Notes 評分備註

• problem 1 is reflection-based — check for genuine engagement with the visualizer, not just generic definitions. HW1 為心得型——確認同學確實操作過視覺化工具，而非僅抄定義。
• problem 2, 3 require working code — verify plots are generated from their own implementation, not copied. HW2/HW3 需繳交可執行程式碼——確認圖表由自己實作產生。
• problem 3 requires statistical rigor (50 runs, multiple parameter settings). Check for proper experimental design. HW3 需具備統計嚴謹度（50次試驗、多組參數），檢查實驗設計是否完整。

EX#6 A* scheduling

本次習題基本說明

進階說明

Handouts 課堂講義

LLM 遭遇 多個限制條件的複雜問題 

專注力模型

Comparison of LLMs

Claude Opus 4.6 optimal (visualizing how AI thinks)

Gemini 3.0 Pro 推理，optimal

ChatGPT 5, end results 流程圖  feasible, not optimal，＊

Grok 4 ，end results 流程圖 ＊

Visualization

Decision Tree (share)

location-based time-evolving diagram (space-temporal transformation ), minimalist

animation on map (national & local) share

課堂練習 

Deadline:  Saturday at 23:59 (one more week)

Send all the share links to  me chang212@gmail.com by email with subject EX#6 [your id, your name]

任選一題就可以

 1. 搶救感恩節晚餐大作戰講義 題目 使用ＡＩ推理或用程式計算出最佳計畫， 然後將求解過程視覺化

推理視覺化（動畫）

style 1

style 2

A* 搜尋樹狀圖（動畫）

style 1

style 2

2. 搶救感恩節晚餐大作戰講義 題目 使用ＡＩ推理或用程式計算出最佳計畫， 然後將得出結果視覺化

狀態圖(State Diagram) 

 (interactive timeline) artifact state diagram

 (interactive timeline) artifact map version

state diagram with aligned timeline

看板圖 (Kanban)

 (interactive timeline)

流程圖(Flow chart)

 (interactive timeline) artifact state diagram

timed flowchart with interactive timeline (share)

Logic circuit design

Logic Circuit Design

slides

EX#5 A* circuit optimization

本次作業詳細說明

  課堂練習 

Deadline: This Saturday at 23:59

Send all the share links to  me chang212@gmail.com by email with subject EX#5 [your id, your name]

1. Design a Two-Stage BJT Amplifier according to goals specified using A* (Parameters to optimize RE1, RC1. RE2. RC2) 

A* search for optimization, share  (c_mu=3 pF, c_pi=22 pF)

2. Optimize LM3886 class AB audio amplifier IC using A* (Parameters to optimize: resistors except the load) 

LM3886

A Search Setup*

• State space: 5 components × 5 E12 values each = 3,125 possible states
• g(n): actual steps taken from the start configuration
• h(n): circuit performance cost (the heuristic) — sum of 5 weighted penalties
• f(n) = g + h: A* priority queue ordering

Cost function (5 components, all minimized):

• Gain: quadratic penalty for deviation from 26 dB (20×)
• f_low: penalty if low-frequency cutoff exceeds 10 Hz
• f_high: penalty if high-frequency cutoff falls below 100 kHz
• Bias: penalizes Rb1 ≠ Rb2 (asymmetric thermal tracking)
• Re stability: log-scale penalty away from 0.47 Ω optimum

Weight/ Forms

Are Re1 and Re2 Always Equal?

the architecture inside the LM3886 and TDA7293

optimizer A* (share)

GPU on a silicon die

 GPU

3D visualization of the GB202 fabrication process , GB202 die, process detailed (share)

CNN signal flow on a block level on a die 

EX#4 Search and Visualizaiion

 Handouts

Chain of Thought with Example, Make your AI powerful

Constraints,  Visualization on energy landscape

Search Algorithms: A*, BFS, Dijkstra

River Puzzles  (Pz 2, Pz 20) by Chain of Thought vs. Search Algortihms

BFS/A* Scalability: 8 vs. 20 persons

  課堂練習 

Deadline: This Saturday at 23:59

Send all the share links to  me chang212@gmail.com by email with subject EX#4 [your id, your name]

任選一題

1.   

Solve Pz 2  by A*/BFS

Animate search tree with synced progression (A*/BFS side-by-side)

 2. 

Solve Pz 20 by  A*/BFS 

Animate search tree with synced progression (A*/BFS side-by-side)

3.  Solve Dog Ball Retrieval by A*/BFS 

Animate search tree with synced progression (A*/BFS side-by-side)

Circuit Design and Optimization

 

• 

  
  
  
  Basic Electronics

• Electronics, Example 6 (2-Stage, A*)

• Electronics, Example 7 (AB, A* Design)
• Electronics, Example 8 (diff pair, OP 741)
• 
  

Mathematical Optimization

 Compare GD, NM Simplex, SA and A* (slides)

https://claude.ai/public/artifacts/dd906391-410f-4fab-b4cd-41e397238496

NM Simplex (Nelder-Mead Simplex)

CMA-ES 

slides

EX#3 Chain of Thought

Handouts

Chain of Thought with Example, Make your AI powerful

Constraints,  Visualization on energy landscape

Search Algorithms: A*, BFS, Dijkstra

River Puzzles  (Pz 2, Pz 20) by Chain of Thought vs. Search Algortihms

BFS/A* Scalability: 8 vs. 20 persons

  課堂練習 

Deadline: This Saturday at 23:59

Send all the share links to  me chang212@gmail.com by email with subject EX#3  [your id, your name]

 

1.   

Solve Pz 2 using  reasoning with each step whether correct or incorrect, showing backtracking. 

Animate the reasoning process in the above synced with river crossing scenarios.  

 2. 

Solve Pz 20 using  reasoning with each step whether correct or incorrect, showing backtracking. 

Animate the reasoning process in the above synced with river crossing scenarios.  

3. Solve Dog Ball Retrieval then animate the solution you got.

Failure example puff

 wait for winds

* Analog IC Design (OP741, diff pair, Placement & Routing)

BJT Differential Pair 

schematic

Quadratic  Placement

ILP+ PathFinder +A* Routing

 (share)

Plus Pseudo Multi-layer Router (95% reduction of congestion md) no via report

corrected vias reported

Plus Symmetry Hard Constraint (cong=0 match>97%, artifact)

neat lohas style

2-stage diff pair (share from very simple diff pair), block diagram,  die, 

schematic (B&W), schematic (color)

corrected

P&R, P&R Best from 1 stage diff pair

tuning (slides)

Phantom exclusion 

artifact

 viz

add overlap penalty in placement stage

benchmark

Catch up 3/11/2026

 Claude 知識遭遇蒸餾事件

EX#2 Solving Electronic Circuits

  AI for solving amplifier electronics (課堂講義）

 建議工具

使用 Claude Sonnet 4.6

使用 ChatGPT 5

使用 Gemini 3 Pro 免費額度最高 1M tokens

使用 Grok 4

Content share

• share only link, pure text, markdown (md)
• no attachments accepted, no html, screen dump, or png
• non-compliant homework will be rejected and returned to you

 課堂練習 

Deadline: This Saturday at 23:59

Send all the share links to  me chang212@gmail.com by email with subject EX#2  [your id, your name]

How to publish a Claude artifact, share alone mayl not work. why?

How to share a ChatGPT link

How to share Gemini Link

四題任選兩題

1. Make the Bode plot of Circuit in Example 1, as below. You must verify the results for correctness.

The results must be scientifically accurate.

Topology B in Schematic

Reference materials:

You may want to generate Python for the Bode plots. Then you  carry it out to the end.

How to make Bode plots 

Make Bode plot:

• Use of Gemini 3.0 Pro slides (Example 1, CE)
• Use of Claude Sonnet 4.6 slides (Example 1, CE)
• 
  

ChatGPT can generate Python like Claude/Gemini, Therefore, you can study  

• Use of Gemini 3.0 Pro slides (Example 1, CE)

Key takeaways:

 

Put your Python to run on Google's colab platform. 

Plots can be made more professional and well suited for formal documents such as reports or academic journals.

You may want to label it more clearly following IEEE standard.

• how to improve Bode plot quality 
• output

2.  Lab Activity: Amplifier Gain & Frequency Response Using AI-Assisted Analysis 

schematic

beta=100, c_pi=100 pF, c_mu=10 pF 

 

beta=150, c_pi=20 pF, c_mu=2 pF Bode Plot (BW)

Bode plot (Color)

3. With the following circuit and its Bode plot, as in Example 3

textbook approximations

Think about the prompt used to generate the Bode plot. Why there is level off at high frequencies?

4. Comparing LLMs: Claude 4.6, Gemini 3.0 Pro, ChatGPT 5.0 (三選二) for the following task.

Animate in 3d (in three.js) Yagi-Uda Antenna, AI pipe inspector, or things you choose. （擇一）

The animation must be scientifically accurate. 

Hints

• Reasoning/Think/Extended Thinking Mode of the AI platform you use may be required for all the problems here

不是主修電子工程的同學，可以練習以下問題

1.

share Claude

Grok 

ChatGPT

ChatGPT 5 eats it alive.

2. Comparing LLMs: Claude 4.6, Gemini 3.0 Pro, ChatGPT 5.0 (三選二) for the following task.

Animate in 3d (in three.js) Yagi-Uda Antenna, AI pipe inspector, or things you choose. （擇一）

The animation must be scientifically accurate. 

Hints

• Reasoning/Think/Extended Thinking Mode of the AI platform you use may be required for all the problems here

3.

source

Claude man in the loop

Gemini 2.5 Pro got it right twice in a row