Day 3 Exercises

Practice tool selection, environmental impact measurement, and green AI strategies.

---

Exercise 1: NLM vs LLM Decision Making

For each task below, decide the most appropriate tool and justify your choice based on technical requirements and environmental impact.

Task A: Patent Classification

Scenario: Classify 200 patents into 5 categories (polymers, ceramics, 
metals, composites, coatings) based on title and abstract.

Output needed: CSV file with columns [Patent Number, Category, 
Confidence Score]

Your decision: - [ ] NLM (BERT-based classifier) - [ ] Small LLM (Llama 8B) - [ ] Large LLM (GPT-4)

Justification: _____

Environmental impact comparison:

Tool	Energy (kWh)	CO₂ (g)	Water (mL)	Time
NLM	___	___	___	___
Small LLM	___	___	___	___
Large LLM	___	___	___	___

Time: 10 minutes

---

Task B: Research Proposal Review

Scenario: Review 3-page research proposal for scientific merit, 
identify gaps in methodology, and provide strategic recommendations 
for improvement.

Output needed: 1-page review with scored rubric and narrative feedback

Your decision: - [ ] NLM (BERT-based classifier) - [ ] Small LLM (Llama 8B) - [ ] Large LLM (GPT-4)

Justification: _____

---

Task C: Chemical Entity Extraction

Scenario: Extract all chemical compound names (with CAS numbers 
if mentioned) from 100 synthesis protocols.

Output needed: List of unique compounds with frequency count

Your decision: - [ ] NLM (ChemBERT) - [ ] Small LLM (Llama 8B) - [ ] Large LLM (GPT-4)

Justification: _____

---

Task D: Strategic Competitive Analysis

Scenario: Analyse 15 competitor patents and 10 recent publications 
to identify gaps in market and recommend 3 strategic R&D directions 
for next 2 years.

Output needed: 3-page strategic memo for executive team

Your decision:

• NLM (BERT-based)
• Small LLM (Llama 8B)
• Medium LLM (Llama 70B)
• Large LLM (GPT-4)

Justification: _____

---

Task E: Document Screening

Scenario: Screen 500 papers to find those discussing "electrospinning" 
and "non-oxide ceramics". Create shortlist of relevant papers.

Output needed: CSV with [DOI, Relevance Score, Brief Note]

Your decision:

• NLM (SciBERT classifier)
• Small LLM (Llama 8B)
• Large LLM (GPT-4)

Justification: _____

Expected efficiency gain vs. using large LLM: ___%

---

Exercise 2: Carbon Footprint Calculation

Part A: Track Your Usage

Track AI usage for the past week (or estimate):

Day	Queries	Avg Tokens/Query	Total Tokens	Task Types
Mon	___	___	___	___
Tue	___	___	___	___
Wed	___	___	___	___
Thu	___	___	___	___
Fri	___	___	___	___
Total	___	___	___	___

---

Part B: Calculate Impact

Weekly footprint: - Queries: ___ - Tokens: ___ - CO₂: Tokens × 0.005g = ___ g - Water: Tokens × 0.1mL = ___ mL

Annual projection: - Queries/year: ___ × 52 = ___ - Tokens/year: ___ × 52 = ___ - CO₂: ___ g × 52 = ___ kg - Water: ___ mL × 52 = ___ litres

Contextualise your impact: - Equivalent driving: ___ kg CO₂ ÷ 0.15 kg/km = ___ km - Equivalent to ___ London-Paris flights (250 kg CO₂ each) - Percentage of UK average (5,500 kg CO₂/year): ___%

---

Part C: Identify High-Impact Queries

Review your past week. Identify:

1. Most token-intensive query: __
   Could it be optimized? __

2. Most frequently repeated query type: __
   Could you create a template? __

3. Query that could use NLM instead of LLM: __
   Expected savings: %

---

Part D: Set Reduction Goals

Based on your footprint, set realistic goals:

This month: - Target query reduction: % - Target token reduction: % - Strategies to implement: 1. __ 2. __ 3. _____

Expected annual savings: - CO₂: ___ kg (% reduction) - Water: ___ litres (% reduction)

Time: 20 minutes

---

Exercise 3: The Green Optimization Challenge

THE BIG CHALLENGE!

The Inefficient Chain

Your colleague wants to understand PLA/graphene electrospinning for a new project. They currently use this approach:

Query 1: "What is polylactic acid?"
Query 2: "What are the properties of PLA?"
Query 3: "What is graphene?"
Query 4: "What is graphene oxide?"
Query 5: "How do graphene and graphene oxide differ?"
Query 6: "What is electrospinning?"
Query 7: "How does electrospinning work for polymers?"
Query 8: "Can you add nanofillers during electrospinning?"
Query 9: "What are the challenges of dispersing graphene in PLA?"
Query 10: "What are typical electrospinning parameters for PLA?"
Query 11: "How does graphene content affect fiber properties?"
Query 12: "What are best practices for PLA/graphene electrospinning?"

Current metrics: - Total queries: 12 - Estimated tokens: ~6,000 - Time: 15-20 minutes (with reading between queries) - Water: ~300 mL - CO₂: ~30 g

---

Your Task

Refactor into 1-2 optimised queries that achieve the same learning outcome.

Requirements:

1. ✅ Cover all knowledge areas from original 12 queries
2. ✅ Appropriate for materials scientist audience
3. ✅ Structured output that's easy to reference
4. ✅ Use CO-STAR or AUTOMAT framework
5. ✅ Specify output format clearly

---

Your Optimised Solution

Query 1 (or only query):

[WRITE YOUR OPTIMISED PROMPT HERE]

Query 2 (if needed):

[WRITE SECOND PROMPT IF NECESSARY]

---

Test and Evaluate

1. Execute in sandbox and time the interaction
2. Compare output quality to original chain
3. Calculate metrics:

Metric	Original	Optimised	Reduction
Queries	12	___	__%
Tokens	~6,000	___	__%
Time	15-20 min	___ min	__%
Water	300 mL	___ mL	__%
CO₂	30 g	___ g	__%

Quality assessment: - [ ] Covers all original topics - [ ] Output is well-structured - [ ] Easy to reference later - [ ] Suitable for materials scientist

Time: 25 minutes

---

Exercise 4: Batch Processing Design

Scenario

You have 20 synthesis protocols to format into standardized templates.

Inefficient approach (baseline):

20 separate queries: "Format this protocol into standard template"
- Queries: 20
- Tokens: ~10,000
- Time: 60 minutes

---

Design Batched Approach

Your batched prompt:

[DESIGN YOUR PROMPT TO HANDLE ALL 20 PROTOCOLS AT ONCE]

Expected metrics: - Queries: ___ - Tokens: ___ - Time: ___ - Reduction: ___%

---

Test Your Approach

Potential issues: - Context window limits? - Quality degradation? - Error handling?

Your mitigation strategies: 1. __ 2. __ 3. _____

Time: 15 minutes

---

Exercise 5: Template Creation

Task

Create a reusable prompt template for a task you perform >5 times/month.

Template name: _____

Task description: _____

Frequency: ___ times/month

---

Your Template

[PERSONA]
Act as {ROLE} with expertise in {DOMAIN}.

[AUDIENCE]
Audience: {TARGET_AUDIENCE}

[TASK]
Task: {SPECIFIC_TASK}

[OUTPUT]
Output format:
{FORMAT_SPECIFICATION}

[METHOD]
Method: {APPROACH}

[CONSTRAINTS]
Constraints:
- {CONSTRAINT_1}
- {CONSTRAINT_2}
- {CONSTRAINT_3}

[TONE]
Tone: {TONE}

[INPUT]
{DATA_PLACEHOLDER}

---

Usage Example

Fill in variables for one actual use case:

[SHOW YOUR TEMPLATE WITH REAL VALUES]

---

Impact Calculation

Before template (writing custom prompt each time): - Time per prompt: ___ minutes - Queries per use: ___ - Monthly time: ___ minutes

After template (fill-in-blanks): - Time per prompt: ___ minutes - Queries per use: ___ - Monthly time: ___ minutes

Savings: - Time saved: ___ minutes/month - Token savings: ___% - CO₂ saved: ___ g/month

Time: 20 minutes

---

Reflection Questions

After completing exercises, consider:

1. Tool Selection: How many tasks could use lighter-weight tools (NLMs instead of LLMs)?
2. Footprint Awareness: Were you surprised by your carbon/water footprint?
3. Optimization Potential: What's your realistic reduction target?
4. Practical Adoption: Which strategies will you implement immediately?
5. Team Impact: How can your team collectively reduce footprint?

---

Action Plan

Based on exercises, create your personal green AI action plan:

This week: 1. __ 2. __ 3. _____

This month: 1. __ 2. __ 3. _____

Accountability: Share plan with a colleague for mutual support

---

Next: Day 4 Overview: Advanced Optimization & Ethics →